Building agentic AI
systems that reason,
not just retrieve.

Built ML pipelines to prepare and process data for model training, integrated into existing backend systems — improving analysis accuracy and reporting efficiency.

Engineered backend systems and data pipelines powering internal reporting tools, with data validation and front-end integration work.

Developed and validated Python-MySQL modules to support internal systems. Built front-end interfaces for internal tools.

University of Warwick, WMG · September 2025 – September 2026

Gujarat Technological University · August 2021 – May 2025

Most LLM equity tools surface correlated market signals without modeling why they move together. SemiBot instead traces an 18-factor causal chain connecting macro conditions to sector outcomes for semiconductor equities specifically, with per-factor attribution rather than a single opaque verdict.

426 observations, 48 tickers: accuracy scales from 84.0% (>$1T) to 41.7% (mid-cap) — confirmed via logistic regression (p<0.001) and robust to sub-sector controls. Mega-cap subset exceeds every comparable published benchmark.

A nine-factor safe-haven analyzer for gold. Rather than assuming the system works and reporting a single accuracy figure, GoldBot is stress-tested across four structurally distinct historical regimes specifically to separate genuine event discrimination from lucky trend alignment.

Diagnostic finding: across 18 distinct events — a war outbreak, a ceasefire, a historic Fed vote — output stayed within a 37–50% band (SD=4.2). Cross-regime testing exposed trend-matching rather than true discrimination, a failure mode a single backtest would have missed entirely.

Gold's monetary-driver explanation is documented in the literature not to hold for silver, whose price is pulled by industrial demand as much as safe-haven sentiment. SilverBot decomposes this into independent industrial and monetary strength scores rather than forcing a single gold-style score.

Industrial signals beat Monetary in 3 of 4 regimes (61.7% vs. 40.0%, p=0.041 in the largest sample) — but the pattern cleanly reverses during the 2020 liquidity crisis (Monetary 95.0% vs. Industrial 65.0%), refining the classical gold/silver duality literature.

A six-signal ensemble for LLM hallucination detection — NLI, retrieval grounding, a three-model LLM-judge panel, and cross-sample self-consistency, run as a parallel LangGraph pipeline. Evaluated with formal significance testing rather than accuracy alone.

72.4–92.2% accuracy across three benchmarks. A quadrant analysis overturned the system's own design assumption: "flagged + consistent" claims were 100% real hallucinations (n=4), beating the architecture's own "high-confidence" category (66.7%).

A human-in-the-loop email drafting agent that never sends without approval. A binary Response/Ignore triage stage filters incoming mail before drafting begins, with full multi-turn interrupt/resume via LangGraph, dual Gmail OAuth, and a Manifest V3 Chrome extension for inline drafting alongside a standalone Streamlit interface.

Zero secrets exposed across the full repo history, verified — two parallel interfaces (v1 Streamlit, v2 Chrome extension with version-history tabs) shipped and maintained side by side.

Full-coverage discovery, research, comparison, and portfolio analysis across all 503 S&P 500 constituents, spanning six pages from dashboard to guided learning. Built on one governing principle: the LLM narrates what the data shows — it never tells the user what to buy.

Passed a 100-question adversarial safety test, 100/100 — results committed to the repo. Implements 9 of 10 recognized agentic-AI design patterns, with the omission (MCP) explicitly justified rather than silently skipped.

Reviewing how AI features are used across conventional and emerging LCA software tools, and how these map onto stages of the LCA workflow.

Unsupervised learning on the ecoinvent 3.11 dataset (25,412 activity-geography records). Identified 49.6% redundancy across 68 raw climate indicators and reduced to 7 conceptually distinct features before clustering.

KMeans (k=8): silhouette 0.689, up from 0.533 on raw features. Isolation Forest flagged 509 anomalies (2%), 57.7% corroborated by clustering.

Compared GPT-4o-mini and Claude Haiku 4.5 on two LCA-specific tasks via direct API calls — structured field extraction, and ISO 14044 goal-and-scope drafting. Claude's extraction errors were zero; GPT-4o-mini's were concentrated in a single, systematic unit-parsing pattern. On drafting, Claude consistently used more precise regulatory terminology (e.g. "comparative assertions," explicit allocation methodology) that GPT-4o-mini omitted in every case.

Claude Haiku 4.5: 100% exact-match vs. GPT-4o-mini: 87.0% (n=100, McNemar's p=0.0002).