AI/ML Researcher & Engineer · Bengaluru
Aryan Haritsa.
I work at the boundary between what models can do and what we can reliably deploy.
Currently: mechanistic interpretability of in-context learning, a 6-encoder Pareto study of dense retrieval architectures, and the occasional 3-week-sprint Kaggle stack.
Now
Apr 2026-
Project Rick @ Kudelski
Agentic chatbot orchestrating AWS Bedrock, Athena, Glue, Teams.
-
Vitruvius @ 90%
Position-sensitivity sweep landed. LSTM upset the hypothesis.
-
Orpheus — mesa-optimization probe
Training small transformers; TransformerLens harness live.
-
ARC AGI 3 · Kaggle
Active submission loop.
01 · Framing
Engineering under
constraint.
Curiosity beneath.
I build systems that have to actually run — a 116-billion-parameter MoE on two consumer GPUs, a 6-encoder retrieval Pareto on a real budget, a ZK-backed KYC pipeline with a cryptographic reason for being.
Underneath the shipping is a quieter question: what are these networks actually computing? Most of my research time goes there — activation patching, head ablations, position shuffles, failure taxonomies. The job of the engineering is to earn the right to ask.
02 · Research threads
Two questions I can't
stop asking.
Orpheus
Mesa-Optimization Probe
Do small transformers trained on in-context linear regression implement gradient descent internally — or merely retrieve from a lookup table? Reverse-engineering the circuits that decide.
Vitruvius
Encoder Archaeology
A 6-encoder Pareto comparison for dense retrieval: transformer vs SSM vs recurrent vs convolutional. Latency. Pruning. Position shuffles. Measured, not argued.
04 · Other work
Engineering & competitions.
TITAN
KYC without exposing KYC. AI document forensics + ZK identity attestation + GNN-based AML transaction analysis — three layers, one structural guarantee.
ARC AGI 3
Abstract reasoning benchmark where brute force stops working.
WaySafe
Safety-weighted tourist route optimization. MIT Bangalore · Sep 2025.
CrediCrew
Semantic resume-to-JD matching. Sentence-transformer embeddings, FAISS recall, cross-encoder rerank. FastAPI + React, sub-second responses.
"The score tells you where you are. The failures tell you where to go next."
— a reflection from Project Ramanujan
05 · Findings feed
Fresh from the bench.
Measurements from the last few weeks of the Vitruvius sweep — five findings that each killed or adjusted a prior belief.
LSTM kept 83% of retrieval quality on shuffled docs.
Predicted sensitivity: ~0.8+. Measured: 0.166 — the lowest of all four architecture families.
Mamba is dominated. BERT-base is dominated.
Pareto-optimal subset on the 3-BEIR avg: gte-small, minilm-l6-v2, lstm-retriever, conv-retriever.
Most heads weren't earning their cost.
Cumulative pruning of MiniLM / BERT / GTE: the majority of attention heads can be zeroed at <5% nDCG@10 loss, zero-shot.
FiQA is the position-sensitivity amplifier.
All 6 encoders peak their shuffle-sensitivity on FiQA (0.24–0.49). NFCorpus & SciFact stay ≤0.27 everywhere.
Transformers 5.x silently drops head_mask.
A 3-hour pod run produced all-zero deltas. One-cell sanity check is now phase-0 for every sweep.
Route-hard-queries-to-big-model doesn't hold.
Unique wins: 127 transformer vs 19 from-scratch across 3 datasets. The partitioning hypothesis isn't supported.
06 · Get in touch
Curious minds
welcomed.
Open to research collaborations, internship conversations (MSR, DeepMind, Anthropic-adjacent), and anyone wanting to compare notes on retrieval architectures or mechanistic interpretability.