AI Researcher at Actioneer · IIT Bombay graduate (B.Tech Mechanical Engineering, minor in Computer Science)
I build the infrastructure around LLM post-training: reward functions grounded in execution, rigorous baselines and ablations, and pre-registered go/no-go experiments. Two threads right now — relational foundation models (learning over a database schema instead of hand-built features) and speech (code-switched ASR, on-device inference).
Across text-to-SQL RL and Hindi–English ASR fine-tuning, the hard problem has consistently turned out to be data and evaluation, not architecture.
📍 Portfolio and writing → jarus77.github.io
| SCRIBE — a Spec → Executor → Planner agent for data analysis across PostgreSQL, SQLite, DuckDB and MongoDB | #1 on DataAgentBench, 81.85% stratified pass@1, verified via a merged upstream PR · code |
| Full-parameter RL for text-to-SQL on BIRD-bench — GRPO plus a custom max-entropy difficulty-weighting estimator, FSDP with CPU offload to fit 7B RL on one node | Reproduced OmniSQL-7B's published baseline within 0.5pp, RL model at 68.12% maj@8. Two research bets killed on pre-registered evidence · code |
| Srota ASR — open-source code-switched Hindi–English speech recognition, fine-tuned from Qwen3-ASR-0.6B | A 3-variant ablation produced a 21.79pp swing from training-data composition alone; script steering gave a 36% relative WER reduction · models · code |
| Srota — privacy-first macOS voice dictation, built solo in Swift/SwiftUI/AppKit | 100+ active users, fully on-device at sub-130 ms end to end. No audio ever leaves the machine |
| Markdrop — PDF to Markdown/HTML with tables, figures and LLM-written descriptions | 210+ stars, 25,000+ downloads. Featured in Japanese tech press and AINews |
| KuralGPT — built with BharatGen | thirukural.ai |
- SynthPID: P&ID Digitization from Topology-Preserving Synthetic Data — CVPR 2026 Workshop (AI4RWC). 63.8±3.1% edge mAP with zero real training data, within 8pp of the real-data oracle.
- Speech-Synchronized Whiteboard Generation via VLM-Driven Structured Drawing Representations — CVPR 2026 Workshop (CV4Edu).
- Federated Cross-Modal Style-Aware Prompt Generation — AAAI 2026 Workshop (SAPP). SOTA on 9 benchmarks (+4.3% unseen classes).
I work through papers by rebuilding every idea in them from scratch. That is my test for whether I actually understood something.
- RelGNN: reading a database as a set of atomic routes
- The Relational Transformer: what survives when the schema changes
- The Jacobian lens: reading the thoughts a model hasn't said
Co-founded Latent Space, an eight-person AI lab at IIT Bombay (CVPR and AAAI publications within a year of founding) · Entrepreneur First founder cohort, Bengaluru · 1st nationwide, Smart India Hackathon 2024 (real-time speech ↔ Indian Sign Language) · 4th worldwide, ICRA 2024 RoboDrive Challenge · Undergraduate Research Award (URA01), IIT Bombay
Most of my code lives with companies, so this account is quiet. Research and product repos are split across @Jarus77 and @Suraj-gameramp.
📫 surajprasad8977@gmail.com · LinkedIn · Google Scholar · X
Always learning. Always building. Open to collaborations.