Oleh D.

Project overview:

The project involved designing and shipping FinWhale, an AI-native financial research platform that fuses a multi-tool conversational agent with structured market-data tools. The platform combines a streaming agent chat backed by LangGraph Cloud, a tabular Data Lens heatmap for cross-ticker quantitative analysis, candlestick Charts with bidirectional daily/intraday linking, and Custom ETF construction. The agent autonomously performs SEC EDGAR filing extraction, SQL queries against years of OHLC market history, web research, and Python execution inside isolated E2B sandboxes - turning hours of manual research into a single conversation.

Responsibilities:

• Architected a multi-tool LangGraph Cloud agent with specialized capabilities for SEC filing extraction, SQL over TimescaleDB, web research, and sandboxed Python — all surfaced to the user as a real-time execution trail.
• Designed the TimescaleDB market warehouse with split-adjusted daily and intraday continuous aggregates, and enforced read-only schema-scoped access so the SQL agent can never mutate production data.
• Built the E2B sandbox layer with a custom Python template so each chat thread gets an isolated, package-preloaded execution environment with file uploads, artifact persistence, and best-effort cleanup on deletion.
• Implemented the Data Lens — a tabular heatmap with per-ticker filters, computed columns, and cross-ticker correlation analysis — wired into the agent via "Open in Data Lens" from any thread's detected tickers.
• Built the streaming chat end-to-end: SSE proxy on AWS EC2, inline ticker extraction and search indexing, citations with hover tooltips and a Sources panel, agent interrupt, and Supabase Auth with magic-link onboarding.

Achievements:

Took FinWhale from spec to a production-grade V1 covering authentication, agent chat, code execution, agent memory, Data Lens, Charts, custom ETFs, cross-surface symbol navigation, and search across chat history. The agent operates on a TimescaleDB warehouse of millions of split-adjusted OHLC bars with sub-second SQL responses, while the chat surface streams tokens, tool calls, and citations in real time through a backend proxy on AWS EC2. Per-user usage tracking and an LLM fallback chain (Claude Sonnet → Gemini → GPT) keep cost and reliability predictable under a $200/month per-user token cap.

Technology stack:

Project overview:

The project involved adapting FinWhale's multi-tool research agent into a vertical product for sell-side M&A advisors. Given a target company preparing for sale, the system autonomously maps the full buyer landscape - strategic acquirers and financial sponsors - scores each candidate on strategic fit, M&A capacity, and recency of activity, and generates ready-to-use outreach hooks tailored to each acquirer's stated strategy. Every claim and hook is grounded in citations from SEC filings, earnings transcripts, deal databases, and public news so the banker can paste a first-touch email into Outlook in minutes rather than spending two to three weeks staffing junior analysts.

Responsibilities:

• Designed a multi-agent pipeline - sector mapper, acquirer ranker, capacity analyst, hook generator, citation validator - orchestrated through a Supervisor agent so each stage hands typed, evidence-bearing outputs to the next.
• Built the buyer-fit scoring model combining strategic fit from 10-K/10-Q text, capital capacity from balance-sheet data, M&A recency, and public-statement signals into a single rankable score with explainable per-dimension breakdowns.
• Engineered the outreach-hook generator: for each acquirer, the agent extracts a strategic gap or stated intent, then drafts a two-to-three sentence opener tying the seller's differentiator to a specific cited acquirer signal - never generic.
• Integrated SEC EDGAR full-text, an M&A deal database, a transcripts/news feed, and PE coverage data into a unified retrieval layer with per-source rate limiting, caching, and freshness tracking so evidence is current at search time.
• Designed the banker-facing UI: target-profile intake, ranked buyer long-list with sortable fit/capacity columns, expandable per-acquirer dossiers, copy-to-clipboard outreach blocks, CSV export, and human-in-the-loop hook editing.

Achievements:

Compressed a typical sell-side buyer-landscape build from 2-3 weeks of associate work into a single afternoon, while producing materially deeper dossiers - each acquirer profile cites the specific 10-K passage, conference remark, or recent transaction that justifies inclusion. Bankers receive a ranked long-list of 50-150 buyers with per-candidate fit scores, capacity-to-pay analysis, prior M&A pattern, and a personalized outreach paragraph keyed to the seller's strategic narrative - eliminating the "generic teaser blast" problem and lifting initial response rates in pilot engagements.

Technology stack: