Bengaluru · memory · graphs · time
I want machines that remember the way people do — not as a pile of similar text, but as facts that begin, hold, and end.
Being
I trained as two things at once: mathematics and mechanical engineering, at BITS Pilani, Hyderabad. One taught me to trust structure. The other taught me that structure has to survive contact with the world. I still think in both languages. When I design a model I am asking whether the math is honest. When I ship a system I am asking whether it will still be true next Tuesday.
I work as a data scientist at SG Analytics in Bengaluru. The job is real: hierarchical Bayesian models for school admissions funnels, multi-agent systems that write and edit decks, sales agents that reason over causal graphs. I like that work. I am not trying to pretend I live in a lab. What changed over the last year is the unit of obsession — from a model that predicts, to a system that remembers, retrieves, and acts. Bit is the other thread: version control for Excel, because a budget should not die as Budget_FINAL_v7.xlsx.
Most AI systems are brilliant amnesiacs. They retrieve a neighborhood of similar text and call it knowing. I want something closer to a life: entities, relations, time. What was true in March. What closed in June. What is still open. That is why I am building mem-void, a temporal knowledge graph for agents, and mini G, a toy DistMult reasoner I refuse to oversell. Honesty about scale is part of the being.
Outside the work I lift, and I eat my way through Bangalore. The body is not a footnote. If I cannot keep a rhythm for myself, I will not keep one for a machine. I prefer execution to performance: say the true size of the thing, ship the next honest version, do not decorate a toy graph as a paper table.
Open edges
Temporal facts. Predicate-aware invalidation. Graph-native retrieval. The opposite of stuffing a context window and hoping.
I want systems that real people touch, under latency and trust constraints. Deeper ownership. Larger blast radius. Less slideshow, more product.
Ablations. Limits. Toy graphs labeled as toy. A statistical engine that stays deterministic even when a language model is invited to explain it.
Language first. Then agents that last. Then, if the memory is real enough, the world. I am not in a hurry to skip the middle.
The public graph
| Predicate | Object | From | To |
|---|---|---|---|
| WORKS_AT | SG Analytics (Straive) Data Scientist. Hierarchical Bayes, LangGraph agents, FastAPI, Azure Databricks. | Jul 2025 | null |
| INTERNED_AT | Harman International DOM Distiller for a browser agent. AutoGen SWARM desktop automation. PPO. | Jan 2025 | Jun 2025 |
| INTERNED_AT | SG Analytics ESG PDF → Excel with Detectron2, DETR, GPT-4o. PPO. | Aug 2024 | Dec 2024 |
| STUDIED | BITS Pilani, Hyderabad M.Sc. Mathematics and B.E. Mechanical Engineering. | 2020 | 2025 |
| RESEARCHED | Tool wear prediction BiLSTM with variational quantum circuits. Prof. K. Kumar. | Jan 2024 | May 2024 |
| BUILDING | Bit Version control for Excel. M1 is the .xlsx round-trip bridge. V0 is two people, one workbook. | Sep 2026 | null |
| BUILDING | mem-void Temporal knowledge graph memory for AI agents. Neo4j. V1. | Jun 2026 | null |
| BUILT | mini G Toy query-conditioned DistMult NBFNet retriever. ~0.45M params. | Aug 2026 | null |
| PUBLISHED | LM-Hypotest Deterministic hypothesis testing on PyPI, with optional LLM interpretation. | Apr 2026 | null |
On the workbench
Version control for Excel — Git-shaped, spreadsheet-native, aimed at FP&A rather than developers. V0 is two people on one workbook: scenarios, review, cell-level merge. What exists now is M1: a round-trip .xlsx bridge so values, formulas, and basic formatting survive parse → write → parse. I do not claim a shipped collaboration product.
Aug 2026A toy query-conditioned DistMult NBFNet retriever, re-derived from G-reasoner. Typed entity–document graph in PyTorch Geometric. I say toy on purpose: ~20 documents, 0.45M params, CPU. On a curated multi-hop set the GNN recovered ranking (multihop MRR 0.14 → 0.82) against the same frozen encoder, with a no-edges ablation so the structure has to earn its keep.
Jun 2026A temporal knowledge graph memory for agents. Raw text becomes episodes, entities, and time-bounded facts in Neo4j. Exclusive predicates close; additive ones coexist. Retrieval is Cypher, not cosine. Inspired by Graphiti. V1 is library-first, no embeddings, no REST. The missing piece I actually want to spend years on.
Apr 2026A Python statistical library on PyPI. The tests are deterministic. Assumptions are checked. A language model may interpret the result, but it is not allowed to become the result. Independent t-test is the current MVP. The philosophy will outlive the test menu.
Query
Not me in the room. A chatbot speaking as me, from a LangGraph backend — opinions included. It should not invent a biography. It should not jailbreak.