Bioinformatics graduate (Loyola University Chicago, 2024). I build scientific software that states only what it can prove.
That sounds like a slogan, so here is what it costs in practice: the projects below publish what was tried and rejected alongside what shipped, name their own weakest evaluation rather than their best, and in two cases document the point where an earlier version of them claimed more than it had measured.
arditmishra.com · LinkedIn · amishra7599@gmail.com
A pre-order check for synthetic DNA. Before an order ships, a separate verifier re-derives the requested edits from the exported sequence and the GRCh38 reference alone and holds the order until it can — the separation is enforced by the function signature, not by convention. Attacked with 22 mutation operators whose ground truth is computed from the delivered file rather than from the checker: 19,823 attempts, 0 false pass, 0 false hold, with two deliberately broken builds shipped so the number stays falsifiable.
The other half is an ADMET property layer over seven endpoints on scaffold splits. DECISIONS.md records what was trained and did not ship, and where the evidence is thin.
A build system for software that builds software. A job arrives as a dependency graph of smaller steps, and each step goes to the cheapest lane that can finish it — local model, free lane, paid only when the work needs it. A step is accepted only when its verification command exits 0, never because an agent reported success. One free model deleted 387 of 389 lines and reported success; that is why the gate exists.
A gene and variant explainer where no language model writes factual content. Every sentence renders deterministically from typed facts in ClinVar, dbSNP, gnomAD, UniProt, MyGene and MyVariant, bound to the source field it cites, and a fail-closed validator refuses anything it cannot ground. Hallucination is not mitigated here; the generation step that would produce one does not exist.
Peptide–MHC class I binding, scored entirely in the browser — no API, no server-side inference, no sequence leaves the machine. Fifteen approaches were evaluated on one identical peptide-grouped split, including CNN, CNN-BiLSTM, Transformer and frozen ESM-2. None beat gradient-boosted trees, and a 1992 substitution matrix matched a pretrained protein language model on a quarter of the head parameters. The full comparison is in DECISIONS.md.
A sequence workbench that deliberately contains no machine learning. Every operation it performs has an exact answer, so the guarantee is one the others cannot make: the same input always returns the same result. An optional Cython k-mer accelerator is measured at 5.3×–11.0× with the pure-Python fallback always available and the engine actually serving a request reported at runtime.
Undergraduate coursework, Loyola University Chicago, 2023, four-person team. EAE mouse-model transcriptomics investigating Sephin1 — Kallisto pseudoalignment, Sleuth differential expression, WebGestalt enrichment. Where the trajectory started.
Languages Python, R, SQL, TypeScript, JavaScript, Bash, Cython
ML & evaluation PyTorch, scikit-learn, XGBoost, SHAP, ESM-2, Hugging Face · leakage-aware splits (scaffold, peptide-grouped, leave-one-allele-out), ROC-AUC/PR-AUC, calibration, adversarial fuzzing against constructed ground truth
Scientific RDKit, Biopython, Kallisto, Sleuth, sequence alignment, VCF normalisation, Golden Gate assembly, GRCh38, ClinVar, gnomAD, AlphaFold
Systems FastAPI, Next.js, React, Three.js, Node, Docker, GitHub Actions, PostgreSQL, AWS, GCP, Vercel
Also five years trading NQ, ES and Gold futures through proprietary firms under defined drawdown and position-sizing limits — where the habit of writing down what actually happened, rather than what was supposed to, came from.