I'm an AI researcher at Actioneer, where I work on
data agents and voice models. I graduated from
IIT Bombay with a B.Tech in Mechanical Engineering
and a minor in Computer Science, and I co-founded Latent Space, an
eight-person AI lab there working on vision–language and multimodal models.
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 occupy me right now:
relational foundation models, i.e. learning directly over the schema of a
relational database instead of over hand-built features; and speech,
from code-switched ASR to 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.
Jun 2026: Took a 7B model through full-parameter RL end to end for
text-to-SQL on BIRD-bench:
reproduced OmniSQL-7B's published baseline within 0.5pp and reached
68.12% maj@8, and killed two research bets on pre-registered evidence
May 2026: Released
Srota Hinglish ASR,
an open-source code-switched Hindi–English model: three models and a dataset on Hugging Face
(code).
A controlled 3-variant ablation showed a 21.79pp swing from training-data
composition alone. Now past 2,900 downloads
2026: Srota,
my local-first macOS voice dictation app, passed 100+ active users,
fully on-device, sub-130 ms end to end
Dec 2025: Joined Actioneer
and built SCRIBE, an agent architecture that ranked
#1 on DataAgentBench (81.85% stratified pass@1), verified via a
merged upstream PR
A 3-stage agent for data-analysis tasks across four database systems (PostgreSQL, SQLite, DuckDB, MongoDB), with runtime schema discovery instead of pre-built indexes
#1 on DataAgentBench: 81.85% stratified pass@1, verified via a merged upstream PR
GRPO plus a custom max-entropy difficulty-weighting advantage estimator (MGPO), FSDP full-shard training with CPU offload to fit 7B full-parameter RL on a single node, and a timeout-safe SQL correctness oracle validating all 1,534 dev golds before training on any of them
Reproduced OmniSQL-7B's baseline within 0.5pp; RL model reached 68.12% maj@8. Two research directions killed on pre-registered evidence rather than intuition.
Fine-tuned and published an ASR model for Hindi–English code-switched speech from Qwen3-ASR-0.6B; diagnosed and fixed a catastrophic negative-transfer failure through deliberate training-data composition
A 3-variant ablation (HiACC-only / OpenSLR-only / union) produced a 21.79pp swing; steering output to natural Devanagari + Latin script gave a 36% relative WER reduction. 2,900+ downloads on Hugging Face.
A corpus of 665 synthetic P&IDs seeded from real engineering drawings, and a technique for steering synthetic generation to preserve graph topology: 63.8±3.1% edge mAP with zero real training data, within 8pp of the real-data oracle.
24 paired Excalidraw demonstrations with narrated audio, every drawing element annotated with millisecond-precision timestamps; a LoRA-tuned Qwen2-VL-7B predicts stroke sequences aligned to speech.
Suraj Prasad, Navyansh Mahla, Sunny Gupta, Amit Sethi
AAAI 2026 Workshop (SAPP)
Multi-scale feature extraction with style encoding and a channel-wise-attention Injection Block, plus a lightweight cross-attention prompt generator that cuts communication overhead in the federated setting. SOTA on 9 benchmarks (+4.3% unseen classes, +1.8% unseen datasets).
WeVerse: Can Vision–Language Models Follow a Story Across Pages?
Suraj Prasad
Independent research, 2026, unpublished
A 7-task evaluation suite (story generation, panel ordering, VQA, retrieval) over 87 webtoon pages. Claude Sonnet 4.6 scores 71.77/100 on narrative quality, 43 points above the closest prior reported result, while every model tested lands at or near chance on temporal panel ordering.
Real-Time Audio to Indian Sign Language Conversion
Suraj Prasad and team
Smart India Hackathon 2024
1st rank nationwide
A bidirectional real-time system: speech → NMT → 3D avatar gesture rendering, and a reverse sign-to-text path (PoseNet + LSTM), robust across speakers and signing styles under strict latency constraints.
3D Object Detection Under Sensor Failure
Suraj Prasad and team
ICRA 2024: RoboDrive Challenge
4th worldwide
Camera and LiDAR fused in a shared bird's-eye-view space for robustness under sensor failure: 0.30 mAP vs. the 0.25 prior benchmark on nuScenes (1.4M images, 390k LiDAR sweeps).
Software
Srota, a local-first macOS voice
dictation app, built solo in Swift/SwiftUI/AppKit. On-device ASR on Apple Silicon with
speculative decoding and Metal/CoreML inference-path optimization, plus a personalisation
layer that learns your vocabulary without sending audio to the cloud. 100+ active users.
Markdrop, an open-source
PDF-to-Markdown/HTML tool built on Table Transformers, Docling and XRef IDs, with
LLM-generated in-place descriptions for non-textual content. 210+ stars, 25,000+ downloads.
Writing
Mostly notes on what I'm working through: relational data, post-training, evaluation, and
what building things has taught me. Full archive here.