Notation matters: chemistry language models and what is inside them

A benzene ring as the sun over a retro grid horizon
Listen
Share

TL;DR. Ask a language model the water solubility of a molecule written four ways (SMILES, an IUPAC name, an InChI, a SELFIES string) and it will often give four different answers. Across 1,072 molecules, 88% got answers more than 0.5 log S apart, with a mean spread of 1.69 log S; a descriptor model fed the same molecules moved 16× less. Looking inside a small chemistry encoder, the representations of two strings of the same molecule progressively diverge with depth (cosine 0.99 → 0.66). This is a full walk-through of the paper, its numbers, and the caveats.

88.0%of 1,072 molecules disagree by > 0.5 log S across notations
1.69mean spread in log S (95% CI 1.62–1.76)
16×smaller spread for a descriptor RandomForest
0.66mean final-layer cosine of two strings of one molecule

The assumption nobody tests

Every chemistry-LLM pipeline silently assumes that the model reasons about the molecule and that the notation is just packaging. The paper puts the assumption plainly: “If a model truly internalises molecular concepts, its property predictions should be robust to this choice.” One molecule has many valid written forms, and they look nothing alike:

  • SMILES linearises the molecular graph as a character string (and one molecule has many valid SMILES).
  • IUPAC names follow systematic nomenclature rules built from substituent prefixes and locants.
  • InChI is a canonical, layered identifier.
  • SELFIES is a token grammar designed so that every string decodes to a valid molecule.

That is a representation-heterogeneity problem in the plainest sense: the same underlying object, several encodings with very different surface statistics. If the model’s answer changes with the encoding, it is responding to the string, not (only) the chemistry. Earlier work had measured this for SMILES versus IUPAC; two things were missing: no study included SELFIES, and none looked inside a model to see where the notation-specific information lives. This paper does both, in two phases.

Four notations, four answersPredicted solubility for propane and 1,4-difluorobenzene under four notations compared with the measured value.Propanemeasured -1.553SMILES -3IUPAC -1.5InChI -2.3SELFIES -3.21,4-Difluorobenzene-5-4-3-2-1012measured -2.636SMILES -4.5IUPAC -3InChI 1.5SELFIES -1.5predicted log S (log₁₀ mol/L); dashed line = measured value
Figure 1 The whole paper in two rows. Propane (CCC · propane · InChI=1S/C3H8/c1-3-2/h3H2,1-2H3 · [C][C][C]) gets four different answers spread over 1.7 log S. 1,4-difluorobenzene (Fc1ccc(F)cc1) is a worse case: measured log S −2.636, yet the InChI form is predicted at +1.5, i.e. “highly soluble”: a 6.0 spread and a sign flip. Source: github.com/amethystani/notation-matters · results/phase1_full_results.json; paper/figures/toc_entry.tex
MoleculeMeasuredSMILESIUPACInChISELFIESSpread
Propane−1.553−3.0−1.5−2.3−3.21.7
1,4-Difluorobenzene−2.636−4.5−3.0+1.5−1.56.0
p-Cresol (Cc1ccc(O)cc1)−2.313−4.6+1.4−1.2−1.56.0
Three real rows from the deposited results. Predictions are the model’s answers (log S); the bold ones are sign flips. Source: results/phase1_full_results.json

Study design

The two-phase study designPhase 1 queries an LLM with four notations per molecule; Phase 2 compares encoder hidden states of two SMILES strings of the same molecule across layers.PHASE 1 · behaviourPHASE 2 · mechanismSMILESIUPACInChISELFIESLLMQwen3.6 35B-A3Btemp 0log S×4spread = max − min over notations · inconsistent if spread > 0.5SMILES ASMILES B(same molecule, rotated SELFIES)ChemBERTa6 layers · CLS stateat L0 … L6cosine(A,B) per layerpatch B ← A at layer kRidge head → Δ|ŷ_A − ŷ_B|
Figure 2 The design. Phase 1 asks whether notation changes an LLM’s answer; Phase 2 asks where, inside a small SMILES encoder, two strings of one molecule drift apart. Note that Phase 2 uses a different, much smaller model than Phase 1. Source: paper Fig. 1 (redrawn); §2
The paper’s overview figure: Phase 1 behavioural study with a sign-flip bar chart, Phase 2 mechanistic analysis of ChemBERTa layers
The paper’s own overview figure (Fig. 1): the Phase 1 sign-flip bars on the left, the Phase 2 pipeline on the right. In this schematic the highlighted layer is illustrative; the data say the most responsive layers are L3 and L5. Source: paper Fig. 1

Data: ESOL, and what was lost on the way

The benchmark is ESOL (Delaney 2004), the standard water-solubility set from MoleculeNet: 1,128 molecules with a measured log S (log₁₀ mol/L). SMILES come straight from the dataset. IUPAC names and InChIs are fetched from the PubChem REST service, SELFIES are generated with the selfies library. A molecule is kept if at least two notations return a valid answer, which leaves 1,072; the 56 dropped ones are PubChem lookup failures.

From 1,128 molecules to the numbers in the paperA funnel: 1128 ESOL molecules, 1072 retained, per notation valid predictions, 790 complete cases.1,128ESOL molecules loadedMoleculeNet / DeepChem delaney-processed.csv1,072Phase 1 retained56 dropped: PubChem lookup failures (≥ 2 valid notations needed)958SELFIES with a valid answer (fewest)SMILES 969 · IUPAC 1,000 · InChI 993 · SELFIES 958790All four notations validthe “complete-case” set: 73.7% of 1,072
Figure 3 Attrition, step by step. Headline numbers use all 1,072 molecules with at least two valid notations; the matched “complete-case” analysis of 790 molecules shows the effect is stronger, not weaker, when attrition is controlled. Source: github.com/amethystani/notation-matters · paper §2.1.1, §3.1.1; results/phase1_full_stats.json

Phase 1: asking one model four ways

The model, and why only one

The reported model is Qwen3.6 (qwen3.6:35b-a3b), a mixture-of-experts model with 35B total and about 3B active parameters, run locally with Ollama in native precision, temperature 0, chain-of-thought disabled, and a 12-token answer limit. It is a general-purpose rather than chemistry-fine-tuned model on purpose, and the MoE design makes deterministic single-workstation runs practical. Four other models were screened and all failed a validity gate:

ModelFamily / precisionModal fractionPearson rValidity gate
qwen3.6:35b-a3bQwen3 MoE, native21.1%+0.755pass
nous-hermes2:34bLlama, 34 B88.8%+0.197fail
qwen2.5:7bQwen2.5, 7 B42.5%+0.055fail
hauhau-q4Qwen3 MoE, 4-bit——fail (4/40 parseable)
hauhau-iq2Qwen3 MoE, 2-bit30.8%−0.177fail
Five models were screened under identical prompts and decoding; only one was fit to measure notation robustness. The gate (modal fraction ≤ 50% and |Pearson r| ≥ 0.30 on 40 random molecules per notation) exists because a model that answers with the same number every time looks “consistent”, or, with a different constant per notation, falsely “inconsistent”. Source: supplement Table S2

The gate matters more than it looks. Read naively, one rejected model (a 34B Llama variant) “yields 98.9% cross-notation inconsistency and a 0% SELFIES sign-flip rate”, numbers that are pure artefacts of the model answering with a different constant per notation. They are deliberately not reported.

The prompt and the metrics

src/notation_benchmark.py · PROMPT ({rep} ∈ SMILES | IUPAC | InChI | SELFIES)
You are a chemistry expert. Predict the water solubility of the molecule below as logS (log₁₀ molar solubility).

Molecule ({rep}): {mol}

Reply with ONLY one decimal number between -10.0 and 2.0. No text, no units.

Each molecule is queried independently per notation: no history, no few-shot examples, no system prompt. The notation is named in the prompt, so the model always knows what kind of string it is looking at; the divergence therefore comes from notation surface features, not from notation ambiguity.

definitions (paper §2.1.3 and src)
spread(molecule)  = max(predictions) − min(predictions)   over notations with a valid answer (≥ 2)
inconsistent      = spread > 0.5 log S
sign flip         = true log S < −1  AND  predicted log S > 0        (insoluble called highly soluble)
parse             = first number matching -?\d+\.?\d* ; accept only if −10.0 ≤ value ≤ 2.0 ; one retry

Phase 1 results

The headline

Analysis setnInconsistent (spread > 0.5)Mean spread (log S)
All molecules with ≥ 2 valid notations1,07288.0% (943)1.686 [1.618, 1.757]
Three notations only (SMILES, IUPAC, InChI)1,01274.9% (758)1.244 [1.180, 1.305]
Complete cases (all four valid)79092.5%1.788
The headline under three views. The 95% intervals are bootstrap (2,000 resamples). Drop SELFIES and the effect is still three in four molecules; require all four notations and it is stronger. Source: results/phase1_full_stats.json; paper §3.1.1

Across 1,072 molecules the median spread between the highest and lowest of the four answers is 1.3 log S: more than a factor of ten in solubility. Only 45 molecules (4%) received four identical answers. The spread is not a threshold artefact:

Share of molecules called inconsistent, by spread thresholdShare of molecules called inconsistent, by spread thresholdspread > 0.10 log S1013 of 1,07294.5%spread > 0.25 log S1003 of 1,07293.6%spread > 0.50 log S943 of 1,07288.0%spread > 0.75 log S867 of 1,07280.9%spread > 1.00 log S654 of 1,07261.0%spread > 1.50 log S503 of 1,07246.9%spread > 2.00 log S282 of 1,07226.3%
Figure 4 How much does the 0.5 cut-off matter? The headline 88.0% (highlighted) is the paper’s choice; loosen it to 0.1 and 94.5% disagree, tighten it to a full 2.0 log S (a factor of 100 in solubility) and still 26.3% do. Source: github.com/amethystani/notation-matters · supplement Table S4; reproduced from results/phase1_full_results.json

Nor is it an artefact of using a range. Replacing max − min by the standard deviation across notations gives a mean of 0.830 log S, and ranks molecules almost identically (Spearman ρ = 0.987). And it is not an artefact of the fine-grained regression framing: collapse the answer to a binary soluble/insoluble call at log S = −2, and on the 790 complete cases 27.8% of molecules still get different verdicts depending on notation.

A split between the average and the disasters

Notationn validMAESign flipsPearson rSpearman ρModal answer
SMILES9691.1275.4%0.7260.77619.9% say −4.5
IUPAC1,0001.0387.3%0.7740.83113.0% say −4.5
InChI9930.8873.3%0.7530.82920.1% say −4.5
SELFIES9581.1361.7%0.6170.66223.7% say −1.5
Per-notation performance. The LLM answers on a coarse one-decimal grid (only 2.4–7.1% unique values across ~1,000 molecules) yet keeps strong rank correlation with the measured values, so low uniqueness is not degeneracy. Source: paper Table 1 and supplement S2/3
Sign-flip rate by notationSign-flip rate by notationSMILES52 / 9695.4%IUPAC73 / 1,0007.3%InChI33 / 9933.3%SELFIES16 / 9581.7%
Figure 5 Catastrophic errors: the share of insoluble molecules (true log S < −1) that the model calls highly soluble (predicted > 0). SELFIES is safest (1.7% vs 5.4% for SMILES; Fisher odds ratio 3.34, one-sided p = 6.2 × 10⁻⁶); IUPAC is riskiest. Source: github.com/amethystani/notation-matters · paper Table 1; results/phase1_full_stats.json
Mean absolute error by notationMean absolute error by notationSMILES1.127IUPAC1.038InChI0.887SELFIES1.136
Figure 6 Average error tells a different story: InChI is best (0.887 log S, 21% below SMILES), and SELFIES is no better than SMILES. “Which notation is best” depends on whether you care about the average or the disasters: the paper’s “previously unreported split between catastrophic and average prediction error”. Source: github.com/amethystani/notation-matters · paper Table 1

Two different notations win two different games. SELFIES minimises catastrophic errors, plausibly because its validity-by-construction grammar prevents the model from reading syntactically impossible token sequences; the paper states this as a hypothesis it continues to hold, not a result. InChI minimises average error. The benefit is not portable either: on the second dataset that passed the validity gate, FreeSolv hydration free energy, InChI goes from best on ESOL to worst of three (MAE 3.054 vs 2.596 for SMILES), and 81.9% of molecules still exceed a 0.5-unit spread. Lipophilicity, BBBP and BACE failed the validity gate, so nothing is claimed about them.

The control: a model that cannot be fooled by strings

LLM vs descriptor model: how much do the answers move?LLM vs descriptor model: how much do the answers move?LLM · mean spread (log S)Qwen3.6, 4 notations1.686Descriptor RandomForest · mean spread4 string routes → same model0.105
Share of molecules with spread > 0.5 log SShare of molecules with spread > 0.5 log SLLM88.0%Descriptor RandomForest9.8%
Figure 7 The control. A RandomForest on eight RDKit descriptors, fed the same molecules through four string routes, moves 16× less (0.105 vs 1.686 log S): for 79.2% of molecules every route yields a bit-identical fingerprint. It is also far more accurate (5-fold CV MAE 0.237 against 0.887 for the best notation), so the claim is about consistency, not accuracy: inconsistency is “a property of models that consume molecular strings”. Source: paper §3.5 (RandomForest code and outputs are not in the public repo)

Which molecules go wrong

Splitting by structural class (assigned by priority-ordered SMILES patterns) shows where the disasters concentrate.

Sign-flip rate by structural class and notationSign-flip rate by structural class and notationSMILESIUPACInChISELFIESCarboxylic / ester (n=51)8.0%2.0%0.0%0.0%Alcohol / phenol (n=177)8.7%5.9%0.0%4.6%Halogenated (n=63)10.7%36.2%26.7%15.7%Aromatic (n=267)5.4%6.0%4.4%0.4%Aliphatic (n=80)0.0%0.0%7.5%0.0%Heterocyclic (n=427)3.9%6.6%0.5%0.0%Amine (n=7)0.0%0.0%0.0%0.0%
Figure 8 Where the disasters live. Halogenated molecules are the danger zone in every notation, and IUPAC names are catastrophic for them (21 of 58 = 36.2%, against 7.3% overall). SELFIES is not uniformly safest: for halogenated compounds SMILES wins (10.7% vs 15.7%), and for alcohols/phenols InChI is perfect (0 of 166). Source: github.com/amethystani/notation-matters · paper Table 2 (corrected seven-class labels)

The most inconsistent class is carboxylic acids and esters (mean spread 2.15, Kruskal–Wallis H = 42.85, p = 1.2 × 10⁻⁷). A careful follow-up in the paper narrows that: of 123 molecules containing a C(=O)O group anywhere, the effect lives in the 51 ring-free ones (2.15), not those with an aromatic ring (2.02) or heterocycle (1.45). The honest statement is “a property of ring-free carbonyl compounds”, and the paper declines to pin it on notation surface features alone.

A bug found by a referee, and why it matters

Peer review caught a defect in the first version’s class labelling. The halogen test was written as a regular-expression character class, which matches any single letter in the set, including the capital C that appears in nearly every SMILES string:

The regular-expression bug that mislabelled 264 moleculesA character class [FClBrI] matches any single letter among F, C, l, B, r, I, so nearly every molecule was labelled halogenated; the fixed alternation Cl|Br|F|I matches whole element symbols.[FClBrI]a character CLASS: any one letterof F, C, l, B, r, ICl|Br|F|Ian alternation: whole symbols;only real halogens match“C” appears in almost every SMILES, so the class label fired on nearly everything.Halogenated molecules, Phase 1 (n = 1,072)327 (buggy label)63 (corrected)264 of 1,072 labels (24.6%) changed. LLM predictions did not: only the story built on the labels did.
Figure 9 A referee-found labelling bug, and what it did to the science. In the first version the 327 “halogenated” molecules included 264 that were not, which masked the real effect; after the fix the paper’s earlier conclusion that SELFIES is safest on halogenated compounds reverses. Source: github.com/amethystani/notation-matters · src/reclassify_results.py; paper §3.2, supplement Table S1

Because the classifier is a priority chain, only the labels after the halogen rule moved, and nothing about the model’s predictions changed, so the 88% and the 1.69 are untouched. What changed is the interpretation: with 264 non-halogenated molecules sitting in the “halogenated” bucket, the real danger of halogens was diluted away. With the corrected seven classes, halogenated molecules turn out to be the most sign-flip-prone class overall, and the earlier claim that SELFIES is safest there reverses. The fix is in reclassify_results.py.

Phase 2: where do the strings drift apart?

Phase 1 shows the symptom in a large LLM. Phase 2 looks for the mechanism in something small enough to open up: ChemBERTa, a 6-layer, 768-wide RoBERTa encoder trained on ZINC SMILES. Being SMILES-only, it cannot be fed four notations, so the paper builds pairs of two SMILES strings of one molecule by rotating a SELFIES token sequence by one position and decoding back, and reports that the retained pairs were checked to encode the same molecular graph (1,051 valid pairs from 1,128 attempts).

Mean CLS cosine similarity between two SMILES of the same molecule, by encoder depthMean CLS cosine similarity between two SMILES of the same molecule, by encoder depth0.600.700.810.921.02L0L1L2L3L4L5L6mean cosine (n = 1,051 pairs)divergence threshold 0.85hidden-state position: L0 = embeddings, L6 = final layer
Figure 10 Divergence builds with depth rather than appearing at any single layer: 0.987 at L1 falling to 0.660 at the final layer. 663 of 1,051 pairs cross below 0.85, with onset concentrated at L2 (186 pairs) and L3 (303 pairs). (L0 is 1.000 by construction: the CLS state before attention depends only on token and position embeddings.) Source: github.com/amethystani/notation-matters · paper Fig. 2 and §3.7; values from paper/figures/figure2.tex

For each pair, the CLS-token hidden state is read at all seven positions (the embedding output plus six layers) and compared by plain cosine similarity (not CKA, not probes). The curve falls steadily: no single layer is responsible. The largest drops are L1→L3 and the final layer. At the final layer, the cosine is related to the downstream disagreement: pairs with lower cosine have bigger prediction gaps (Spearman ρ = −0.371, p = 1.4 × 10⁻³⁵), and the mean gap rises from 0.324 log S in the most similar fifth of pairs to 1.129 in the least similar fifth.

Final-layer cosine similarity by structural classFinal-layer cosine similarity by structural classAromatic293 pairs · mean gap 1.120 log S0.417Heterocyclic380 pairs · gap 1.0290.573Aliphatic74 pairs · gap 0.5500.899Alcohol173 pairs · gap 0.5410.939Carboxylic51 pairs · gap 0.5580.945Halogenated73 pairs · gap 0.4290.953Amine7 pairs · gap 0.6470.961
Figure 11 Which molecules diverge most inside the encoder? Aromatic and heterocyclic ones (cosine 0.42 and 0.57; prediction gaps above 1 log S). A halogen is a single unambiguous token in any string, whereas aromatic rings admit kekulised vs lowercase forms and different ring-closure numbering. The paper offers this as a hypothesis, not an established mechanism. Source: paper Table 4

Can you repair it? A hidden-state replacement test

Divergence is a correlation. To probe whether the diverged state carries the disagreement, the paper runs a replacement experiment, deliberately not called “causal”: during the forward pass of string B, overwrite the CLS state at layer k with string A’s state, then let a Ridge regression head predict log S and see whether the two predictions move together.

paper Eq. 1 (hook-based representation replacement)
# patch layer k of the SMILES_B forward pass with the SMILES_A state, then read out with the Ridge head
Δ(k) = |ŷ_A − ŷ_B|  −  |ŷ_A − ŷ_B^(k)|          # positive = the patch narrowed the gap
Effect of patching one layer’s CLS stateGrouped bars per layer for three controls: patching in the true partner, a random molecule, or reversed direction; the true partner is near zero while random molecules hurt more with depth.-0.8-0.6-0.4-0.2+0.0+0.2-0.006-0.076+0.008L1-0.012-0.223-0.086L2+0.089-0.458-0.092L3+0.038-0.688-0.009L4+0.096-0.826+0.000L5true partner (same molecule, other string)random moleculereversed directionΔ log S
Figure 12 Does swapping in the other string’s hidden state repair the prediction? Only barely. The true partner’s state helps by at most 0.096 log S (L5), about 11% of the 0.873 mean gap, and still makes 42.9% of pairs worse; a random molecule’s state does far worse and increasingly so with depth (−0.826 at L5). Reading: the diverged states are still strongly molecule-specific, “a failure to converge on a common encoding rather than a loss of chemical identity”. Source: supplement Table S3 (positive = gap narrowed)
LayerTrue partnerRandom moleculeReversedCanonical (own state)Pairs made worse
L1−0.006−0.076+0.0080.00049.1%
L2−0.012−0.223−0.0860.00050.0%
L3+0.089−0.458−0.0920.00044.1%
L4+0.038−0.688−0.0090.00045.7%
L5+0.096−0.826+0.0000.00042.9%
The controls behind the intervention. The “canonical” row is exactly zero for every pair: swapping a state with itself changes nothing, which rules out the hook mechanism, CLS substitution as such, and the Ridge head as sources of the effect. The final layer is excluded because the head reads it directly, so patching it just returns the source molecule’s prediction. Source: supplement Table S3
  1. The effect is real but small. Significant at L3 (p = 6.1 × 10⁻⁶) and L5 (p = 4.7 × 10⁻⁵) after Bonferroni correction, marginal at L4, absent at L1 and L2.
  2. It is specific. The true partner beats a random molecule at every layer, in 58.7% of pairs at L1 rising to 75.5% at L5.
  3. It is not an artefact of the machinery. The canonical control is exactly zero for all 1,051 pairs.

Consistency is not accuracy

Finally, are inconsistent predictions just inaccurate ones? Measured per pair, the gap between the two predictions has essentially no relationship with the absolute error (Spearman ρ = −0.059, p = 0.055, n = 1,051). Being consistent and being right are separate properties, which is why fine-tuning for accuracy alone should not be expected to fix notation sensitivity. (A non-significant result is a failure to reject, so the paper’s word “orthogonal” is best read as “no detectable dependence at this sample size”.)

What to do with this

  1. Match the notation to the error you fearSELFIES for fewest catastrophic sign flips (heterocyclic 0/378, carboxylic/ester 0/51, aromatic 1/267); InChI for lowest average error and for alcohols/phenols; SMILES over SELFIES for halogenated compounds; avoid IUPAC names there (36.2% sign flips).
  2. Say which notation you usedIt is a methods-section variable, not a detail. The same model on the same molecule gives different answers.
  3. Do not treat single-notation accuracy as molecular understandingA model can be accurate in one notation and inconsistent across them. Test consistency separately.
  4. Aim for convergence, not agreementCross-notation losses applied across intermediate layers, or a multi-notation encoder, target the representation rather than just matching outputs.

What this paper does not claim

  • That any notation is universally best; the ranking is task-dependent and does not transfer to FreeSolv.
  • That the effect holds for chemistry-fine-tuned models, other LLMs, or classification tasks. One model passed the gate; the others failed it, and no hosted API model was evaluated.
  • That the encoder in Phase 2 is the Phase 1 model; it is a different, much smaller model, and the Phase 2 pairs are within-SMILES rather than true cross-notation pairs.
  • Formal causal mediation. The experiment is a representation-level replacement.
  • That the descriptor model is better overall; it is more accurate, and the argument is only about consistency.

Code map and how to run it

ScriptRole
src/notation_benchmark.pyPhase 1: classify, fetch IUPAC + InChI from PubChem, encode SELFIES, query the LLM in four notations, compute spreads, sign flips, MAE, Fisher and Kruskal–Wallis; checkpoints every 25 molecules
src/reclassify_results.pyRelabel Phase 1 rows with the corrected classifier and recompute class statistics (no new inference)
src/mechanistic_analysis.pyPhase 2: SELFIES-rotation pairs, ChemBERTa CLS states at seven positions, Ridge head, per-pair cosines, hook-based hidden-state replacement
src/model_validity_probe.pyThe validity gate: 40 random molecules × 4 notations; pass iff modal fraction ≤ 50% and |r| ≥ 0.30
src/second_model_benchmark.pySame protocol on another open model, with failure accounting by cause
src/multidataset_benchmark.pyThe protocol on FreeSolv, Lipophilicity, BBBP and BACE
src/build_notation_cache.pyFetch IUPAC/InChI once with polite pacing and backoff so rate limits are not mistaken for model failures
shell
pip install -r requirements.txt     # numpy scipy scikit-learn pandas requests selfies torch transformers rdkit
ollama pull qwen3.6
bash run_benchmark.sh                   # builds the notation cache, then runs Phase 2 and Phase 1

Phase 1 takes roughly 5–8 hours and Phase 2 1–4 hours on a single workstation. The repository deposits the Phase 1 result files; the Phase 2 outputs, the RandomForest baseline code and the second-dataset outputs described in the paper are not in the public repository. Citation: Animesh Mishra, Notation matters: cross-representation inconsistency in chemistry language models and its mechanistic origins, Digital Discovery, 2026, DOI 10.1039/D6DD00309E.

All articles

Research and building cool stuff

© 2026, Animesh Mishra

GitHub|LinkedIn

New Delhi, India