Arjun Desikan

• Built an LLM-driven specification synthesis system spanning C/ACSL and Rust/Lean 4, integrating Frama-C/WP, Alt-Ergo, Charon/Aeneas, mutation-guided refinement, executable counterexamples, and kernel-checked proofs.
• Achieved a 0.976 mean mutation kill rate versus a 0.622 baseline across 189 Lean benchmarks while reducing prover calls 11-fold (19.9 vs. 224.8) and wall time 2.4-fold (71 vs. 168 seconds) with a five-gate validation oracle.
• Closed an end-to-end verification chain on the production Rust crate gcd 2.3.0: proved 11/11 specifications with 0 refuted and generated a 40-theorem, zero-sorry Lean proof comprising 8 soundness goals and 32 supporting lemmas.
• Delivered 10 certified C artifacts with Frama-C/WP and Alt-Ergo, discharging up to 27/27 proof obligations; imported 35 benchmark algorithms and rejected weak specifications through property-based and TCE-filtered mutation testing.

• Built PyTorch multimodal CNN-MLP latent-fusion and DeepONet baselines for visual-to-tactile acceleration prediction; in preliminary runs, the time-domain DeepONet achieved 1.11 test RMSE, 35% below the CNN-MLP baseline's 1.72.
• Evaluated material generalization using exclusive texture splits; used Fourier encodings to encourage oscillatory predictions.
• Implemented a Fourier DeepONet on normalized log-FFT magnitudes to address spectral bias; achieved 0.92 preliminary test RMSE (17% below time-domain DeepONet; 46% below CNN-MLP) and explained over 80% of lower-frequency X-axis variance.

• Trained 4.9M-param financial time series transformer encoder on 24.8M observations (499 tickers, 12 features, 4,149 daily bars).
• Designed temporal/cross-asset architecture extending temporal neighborhood encoding (TNC) with graph convolution layers and forward-aligned objectives over 120-day historical correlations or learned attention; submitted to ICLR 2027.
• Benchmarked 9 representations across 21 targets, 3 horizons, and 3 seeds on 37,500 embargoed windows; at 21 days, achieved +0.110 relative-volatility and +0.090 beta rank-IC lift over strongest baselines with under 0.015 seed variation.
• Developed production features and bug-fixes for Mantis (lab software) using React, TypeScript, FastAPI, Zustand, and Docker.
• Designed full-stack swarm intelligence simulator Mantis feature, modeling evolving agent representations using aligned UMAP.

• Analyzed GISTIC copy-number and RNA-seq data (3,527 TCGA tumors, 12 cancer types) using PCA, ANOVA, MLR, and PLS.
• Directly associated SCNA burden with transcriptomic effective rank across quartiles (1000+ bootstrap resamples); PLS linked intracellular transport and immune response to burden with R-squared 0.883; abstract published in STEM Fellowship Journal.

• Produced 2021 statistical mortality brief over county population and 5 demographic subgroups; used by county to inform policy.
• Built an end-to-end SQL/R mortality pipeline for age-adjusted rates, t-tests, chi-square and K-S tests, and ggplot2 visualizations.
• Forecast influenza, RSV, and COVID-19 cases (XGBoost, Prophet, LSTM) reducing RMSE by 50% versus ARIMA baseline.

• Built a two-agent harness testing whether a misaligned coding sub-agent pulls its conductor LLM toward reward hacking; diff-of-means activation steering on Qwen3-8B moved shortcut adoption from 22% to 3% or 75% at full coherence.
• Linear directions read corner-cutting (0.935 AUROC), but same-wording controls showed they encode content and negation, not intent (reversed AUROC 0.12–0.39); building a GRPO + LoRA post-training environment to induce reward hacking.

• Team benchmarked 24 tasks/10 sites: naive retrieval cut success from 21% to 13%. Calibrated 384-d MiniLM/pgvector with a strict similarity floor and 0.7 similarity/0.3 confidence ranking, retrieving lessons in 10/14 failure cases and none for controls.
• Built a Next.js run inspector with trace replay, baseline-vs-memory view, and keyless test backend; orchestrated a single-task live-site pilot where one lesson raised same-model success from 2/4 to 6/6, cutting failed actions 69% and task time 57%.
• Live-tested MCP Streamable HTTP with Claude agents; contributed to the Playwright loop and WebArena-verified scoring, using Wilson lower bounds for confidence and Postgres constraints/RLS for idempotent outcome recording.

• Built a 5,300-line JavaScript/Electron app bringing Git to Excel: parses OOXML into cell-level diffs for inputs, formulas, recalculated values, and formatting, with commit/history/restore/revert/branch/stash/push/pull workflows.
• Engineered an OOXML normalizer to suppress save-only diffs: hashes styles by font, fill, border, and number format; expands shared formulas; resolves relationship IDs; and strips volatile metadata and chart caches so unchanged re-saves produce no diff.
• Implemented cell-keyed three-way Git merges with LCS/diff3 for ordered content; auto-merges disjoint edits, ignores formula-cache churn, and safely rejects unsupported files; 39 tests cover Git/workbook cases.

• Modeled daily ICU states with an absorbing Markov chain fit to 14,000+ observed transitions across 186 hospitals; derived closed-form mortality and length-of-stay estimates, matching observed mortality (8.32% vs. 8.31%).
• Built a vectorized Monte Carlo engine for VaR₉₉/expected shortfall; control variates cut estimator variance 2.6x, common random numbers cut scenario variance 2.5–3.7x, and ignoring parameter uncertainty understated VaR₉₉ by 12% at n=50.
• Compared 7 models with patient-level walk-forward/cross-validation and Bayesian shrinkage, reducing out-of-sample log-loss 19% vs. no-change baseline; a bootstrap likelihood-ratio test rejected one-day memory (p=0.002).

• Built a Solidity/React marketplace for speedrun glitches on Base Sepolia testnet with wagmi/viem wallet integration.
• Implemented AES-256-GCM/IPFS content delivery with on-chain commitments, time- and sales-based price decay, optimistic settlement, reputation stakes, and decay-gated jury disputes.
• Wrote 34 Foundry tests and live end-to-end checks covering pricing, escrow settlement, bonded disputes, and juror eligibility.

• Trained a PyTorch 1D CNN for ECG abnormality classification, achieving an AUC of 0.98 in first-pass modeling.
• Used Grad-CAM to explain ECG predictions clinically, identifying decision-driving ST-segment and QRS complex regions.
• Optimized model performance with Optuna hyperparameter search over dropout, batch sizes, and learning rate scheduling.

• Built backtester for EMA crossover and mean reversion; customizable transaction costs, EMA horizons, position sizing, etc.
• Integrated Monte Carlo simulations (block bootstrap, parametric) for risk analysis and stress-testing across any input ticker.

• Built a React Native app and website for skateboarding progression tracking and text-based coaching, with 40+ interconnected tricks visualized in a directed prerequisite graph.
• Implemented a RAG pipeline grounded in skateboarding trick advice from 50+ wiki webpages scraped with BeautifulSoup.
• Prototyped Google MediaPipe/VLM video analysis comparing uploaded tricks to professional examples at 5 technique keyframes.

• Led a 3-person team building personalized learning pathways across 1,000,000+ research papers from natural language queries.
• Designed a content-based recommender with SciBERT embeddings, pgvector similarity search, LLM-based query expansion, and the kneedle algorithm to auto-tune pathway length.

• Lead 4-dev team building an adaptive quizbowl study app (500+ users) across Electron, FastAPI, and PostgreSQL stack.
• Engineered NLP ingestion, topic clustering and production epsilon-greedy softmax recommender (160,000+ questions).
• Prototyped RL recommender (DQN) against production softmax system using simulated student across varied reward structures.

How we serve question recommendations for a trivia vertical where both timing and accuracy matter, when every question comes with only a broad category and a broad difficulty.

An exploration of my relationship with the USA and its founding ideals.

An exploration asking whether naive dreaming serves us better than uncomfortable truth.