Solo ProjectsEverything here was designed, built, and shipped by one person.

Genomic-Foundation-Model RNA Therapeutic Candidate Ranker · Murai Labs

A modular pipeline that ranks RNA therapeutic interventions — antisense oligos, splice-switching oligos, base edits — using an ensemble of genomic foundation models (Evo 2, a 4-replicate Borzoi ensemble, Enformer, ViennaRNA). You bring a gene, an exon, and a tissue; PINNAL tells you which of the ~200–5,000 possible binding sites to wet-test first, and shows its work. It is a candidate ranker — not a target-discovery tool — that closes the loop between computational prediction and wet-lab experimentation.

Personalized AI Storytelling Platform for Children

A personalized children's storytelling platform that generates deeply original, emotionally coherent stories — not generic AI output — using a multi-pass story engine and persistent character memory.

Tamil-First Offline ASHA Triage · Gemma 4 Good Hackathon

A three-tier offline community-health intelligence system for Tamil-speaking ASHA workers in rural India. Tier 1 runs a 5GB Q4_K_M Gemma 4 E4B GGUF entirely on an Android phone — model, IMNCI protocol engine, AES-256 encrypted SQLite — with zero network dependency for a triage decision. Tier 2 is a PHC doctor console; Tier 3 is a district health dashboard with geohash heatmaps and cluster alerts.

Cultural Memory Passport

Cultural Memory Passport for diaspora families using JSON-LD schema, JWT consent, and a Claude-powered story engine.

Governed AI Software Refinery

Governed AI software refinery for enterprise code generation — built for Microsoft AI Dev Days Hackathon. Brings governance and auditability to AI-assisted code generation, not just raw throughput.

Probabilistic Reconstruction of Lost Pallava Temples

A research system for reconstructing the six lost Pallava temples of Mahabalipuram using evidence-tiered inference, Agamic architectural constraints, posterior reweighting, and per-element uncertainty maps. The system does not claim a single historical truth — it characterizes the posterior over plausible reconstructions given the evidence.

AI-Driven Historical World Reconstruction for VR

An AI system that reconstructs historical Madras (Chennai) across time (1800–2000) as explorable VR environments, combining archival data, structured world modeling, and deterministic multimodal generation.

Seam-Masked Latent JEPA for Tamil Akshara Recognition — A Cheap-Baseline Falsification

A concluded negative-results study. It tested whether masking a Tamil akshara at its grapheme seam (the base-consonant / vowel-sign boundary) and predicting the masked region's latent (I-JEPA style) teaches a model the abugida's base × sign composition better than generic masking or pixel reconstruction — and thereby recognises rare, low-frequency compounds more accurately. On rendered multi-font Tamil at evenings-scale (ViT-Tiny/8, ~20k augmented instances, one RTX 5090), the hypothesis was falsified on both of its pre-registered kill criteria. The honest negative is the contribution.

Heterogeneous Distributed LLM Inference · ANE + CUDA + Metal

A distributed inference pipeline that lights up five compute targets inside a single EXO-orchestrated cluster: NVIDIA CUDA (RTX 5090) for prefill, Apple Metal for decode, the Apple Neural Engine for speculative-decode drafts, CPU for KV cache management, and NVMe for Flash-MoE expert streaming. Composes TriAttention KV compression with TurboQuant+ weight compression so a consumer WiFi link stops being the bottleneck.

Ancient Prior Model — Tamil Grammar Priors for Stellar Variability

A transformer pipeline that classifies TESS stellar variability using Tamil-grammar-derived attention biases as a structural prior. The Ancient Prior Model (APM) tests whether canonical grammar tokens, cyclic positional encodings, and co-attestation-based attention bias improve few-shot variability classification on real astrophysical light curves. Targeted at EAS 2026.

Adversarial Multi-Model Deliberation Engine

Adversarial multi-model deliberation engine (Claude / GPT-4o / Gemini / Grok) with Coherence Trap detection and Polymarket integration. Stress-tests AI reasoning through opposed debate rather than fragile single-shot prompting.

Context-Injection & SDLC Orchestration Platform

A developer-first system for injecting structured context into AI workflows across the SDLC — planning, execution, and verification each get their own context layer.

Does Karpathy-Style Code Clarity Improve SWE-Bench Repair?

A research dataset and training pipeline testing whether Karpathy-style code clarity (descriptive names, short functions, minimum abstractions) improves automated software-repair performance on SWE-bench. The seed batch is 201 hand-rolled Karpathy-style examples; the pipeline runs three augmentation passes to reach ~10K rows across splits before LoRA fine-tuning on Qwen2.5-Coder.

Open-Source Household OS for Companion Robots

An open-source operating system for embodied companion robots. The first reference build runs on a Reachy Mini–class desktop robot paired with a local compute host — doing local-first multi-user voice interaction, identity recognition, persistent per-person memory, embodied expression, and clean handoff to agents for cloud-touching tasks. Three replaceable layers: a Hardware Abstraction Layer, Chitti Core (perception, identity, context, conversation, expression, dispatch), and pluggable agent backends behind a stable RPC contract.

Claude-Powered Chrome Extension for Real-Time Debugging

A Chrome extension that watches the browser console in real time — console.error, console.warn, uncaught exceptions, and network failures — and uses Claude to explain each error and propose a code fix, shown in a side panel so it never blocks your work. Local-first and privacy-respecting: no backend, your Anthropic API key stays in Chrome sync storage, and it only monitors a domain whitelist you control.