Prateek Mulye · Software engineer
11+ years across banking, data platforms and manufacturing. I build reliable software and bring that experience to practical AI applications.
Based in Santa Barbara, California. Open to engineering opportunities.
A combined result of partitioning, indexing and replicas; no benchmark is exposed here.
Career account; employment outcomes are not independently reproduced here.
Measured in test environments.
No accuracy or savings claim.
The work, connected
Four areas of work, connected by questions about state, trust and delivery. Choose one to see the decisions behind it.
Career account · Capital One, via Cognizant · 2018–2021
I owned Kafka workflows using Spring Boot and DynamoDB: dead-letter handling, bounded retries with backoff and idempotent replay.
Recovery controls are not a universal exactly-once or zero-data-loss guarantee. The separate OpenResty result of 12K+ transactions per second was measured in test environments.
Career account · HG Insights · 2021–2025
I built an idempotent contract-processing pipeline for 80K+ records, reducing reporting turnaround by 60%. Oban Pro also orchestrated scheduled, on-demand and event-triggered company-spend refresh.
Spend refresh moved from a manual monthly batch to a weekly cadence. My contribution was orchestration; the spend formula was an existing service.
Source: public career account, engineering decisions and résumé in the reviewed public record. Employment outcomes are not independently reproduced here. Read the résumé
Career account · HG Insights · 2021–2025
At more than 30M PostgreSQL rows, I used partitioning, composite indexing and read replicas to reduce dashboard latency by more than 60%.
The decision
Partitioning narrows the relevant data; composite indexes support predicates within it. Read replicas separate read traffic from the primary.
These are general tradeoffs, not a reconstruction of HG Insights' schema or deployment. The reported dashboard result combines the changes; it does not isolate each one's effect.
1 of 8 illustrative partitions is in scope.
WHERE month = 'Aug'With JavaScript off, the latest-month illustration remains visible.
Authored employment evidence. No production query trace or benchmark is exposed here. Read the public résumé
“He scaled PostgreSQL beyond 30M rows using partitioning, composite indexing and read replicas, reducing dashboard latency by more than 60%.”
Public experience record, HG Insights. The geo-standardization service separately normalized 15M+ US and Canada company records into canonical shapes, with a phased rollout led with stakeholders.
Career account · Capital One, via Cognizant
The public résumé places DynamoDB alongside the Kafka recovery workflows I owned. AWS is also part of my broader professional tools record.
The published record does not expose the table design, deployment topology or service-by-service AWS ownership. No broader AWS architecture is inferred here.
Reviewed source · ChatFormula1 v2
My developing v2 connects a Python/LangGraph retrieval pipeline, an Elixir/Phoenix gateway and a typed React streaming client.
Schema-validated routing, parallel retrieval, fallback.
Streaming, conversation persistence and spend accounting.
A reducer handles duplicate and replayed stream events.
The decision
The client can avoid repeating token batches and identify a gap. Those controls make transport behavior inspectable; they cannot establish that an answer is correct.
The v2 graph is stateless. Gateway persistence does not mean Redis graph checkpointing. The older v1 demo and developing v2 are separate. Current v2 deployment and live-model quality are unverified.
Inspect the pinned React reducerChatFormula1 source commit 3574414d2786. The existing portfolio wrapper uses the Apache-2.0 reducer and adds bounded input validation. License
Assay is a separate Python/LangGraph and FastAPI project with an evaluation harness. It originated in SuperDataScience CP044; my contribution is not sole authorship of the upstream project. A harness alone is not a measured quality result. All projects & historical evaluations Assay source
Runnable local demo · synthetic events
Uses the original reducer in your browser. No model, API or database request.
Delivery scenario
Choose a fixture, then step through its events.
A late batch can fill the visible text while the gap stays flagged. The real app requests a persisted message; this demo does not make that request.
Open a work area to explore its decisions and evidence.
Career
My recent applied AI work builds on a longer software engineering career. Earlier banking and data-platform roles are not years of AI engineering.
Senior Software Engineer, Manufacturing. Internal Gemba reporting and action-tracking applications with Power Apps and Power Automate; manufacturing operations, recipe authoring and internal AI enablement.
Broader plant rollout is expressed interest, not a completed deployment. No quantified savings or externally shipped AI product is claimed.
Software Engineer. Data standardization, PostgreSQL performance and Elixir/Oban Pro orchestration.
Software Engineer. Java/Kafka recovery and OpenResty service virtualization.
Personal applications
These separate browser applications pair local model execution with explicit review. Application links are from the public record; availability has not been retested in this offline review.
Exact constraints first, MiniLM ranking second, a reviewer decides.
No operational matching-accuracy claim. Does not move money or post entries.
A TensorFlow.js regressor must beat the better baseline before flagging.
All three public-data trials withheld flags. Assumed occupancy; no savings claim.
Whisper suggests caption differences. A person listens, edits and exports.
Speech can be misheard. Synthetic voice checks are not an accuracy benchmark.
Inputs and review decisions stay in browser memory; export before leaving. Caption Review downloads public model files. These are personal applications, separate from employer work.
Beyond work
I hike when I travel, and I will walk a city end to end before I take a car through it.
Languages English, Hindi, Marathi · German A1–A2, learning.
Travel, photographs and the personal record
Start a conversation
I’m Soren, Prateek’s AI representative. Ask me the questions you’d ask Prateek about his work. I’ll explain what the sources show, share my read, and flag what he’d need to answer himself.
AI voice · ElevenLabs generated it from the introduction above. It plays only when you press the button.
Soren can research approved public sources and write a conversational interpretation grounded in the material it retrieved. Expand a source to inspect that material. Citations help you check an answer; they do not guarantee its accuracy. If conversational generation is unavailable, Soren identifies the evidence fallback. LinkedIn is a profile link, not a connected private account.
Your question and up to two prior visitor questions (6,000 characters total for those prior questions) and up to four public evidence references go to Cloudflare and, when enabled, Workers AI and optional Groq fallback. Previous answers are not sent as follow-up context. Laya may classify the current question. Providers may retain data under their policies; avoid sensitive information. Email and phone redaction does not anonymize text. Approved GitHub and Hugging Face sources may be read without account credentials. Up to six completed answers, your draft and the bounded follow-up context stay in this tab’s sessionStorage while you browse or reload. No transcript is stored in localStorage or an account. Clear conversation deletes this tab’s saved conversation, draft and follow-up context. Browser session restore may retain tab storage after closing; use Clear to remove it explicitly. If tab storage is unavailable, conversation stays in memory. Local brief selections and the email preview stay in memory only.