--- title: "Manas Nikam | Forward Deployed Engineer, remote from Mumbai" canonical: https://manasnikam.dev/ updated: 2026-09-23 --- # I make AI work inside existing systems. I'm Manas Nikam, a forward deployed engineer in Mumbai, remote for a Canadian client since May 2023. 8+ years wiring payments, suppliers and CRMs, now wiring LLM agents into them; my JonView integration took a supplier booking from about 10 minutes to about 30 seconds. - 30+: providers behind one payments interface - 10 min to 30 s: per supplier booking - 3 agents, 3 days: on Cloud Run, 2 over vendor MCP servers Live: https://agentready.prescalesystems.workers.dev (WebMCP readiness benchmark, built and submitted September 2026) ## Case studies Three engagements, three numbers, and what I would do differently on each. The AI I ship goes into systems customers already run: Gemini OCR on a printer-services company's live invoice flow, with a spend cap and a per-invoice cost line, built on the same integration habits I used for 30+ payment providers. Structured outputs, MCP clients, a SHA-256 verifier for model claims, a 95%-confidence-interval eval runner and a per-invoice cost line: the parts around the model are the work. - [Integrating VIA Rail and JonView into a tour operator's booking system](https://manasnikam.dev/case-studies/via-rail-jonview-supplier-integration/): 10 min to 30 s, per supplier booking, JonView. Three travel suppliers wired into a live itinerary system for Fresh Tracks Canada: JonView search and booking with a circuit breaker and state sync, VIA Rail built end to end in five dated steps, a Rocky Mountaineer scraper. Per-product booking time went from about 10 minutes to about 30 seconds. - [Gemini invoice OCR and automated Tally invoicing for a printer-fleet business](https://manasnikam.dev/case-studies/cd-infoware-gemini-ocr-tally-invoicing/): 10 instances, hard cap on Vertex OCR; cost logged per invoice. Live for CD Infoware, an Indian printer-services company: invoice OCR on Gemini through Vertex AI with a 10-instance cap and per-invoice cost in USD and INR, plus a nightly Konica Minolta meter scrape that drafts one Tally-format GST invoice per device per month. 111 commits, all mine. - [Three agents on Cloud Run in three days, two integrated with vendor MCP servers](https://manasnikam.dev/case-studies/adk-agents-mcp-cloud-run/): 3 in 3 days, agents on Cloud Run, 2 over vendor MCP servers. 3 production agents on Cloud Run in 3 days, 2 integrated with vendor MCP servers (ClickHouse, Grafana): a screenplay clearance pipeline, a retention analyst over mcp-clickhouse, and a VFX war-room agent over mcp-grafana that writes back annotations and incidents. Each live with a demo video. Built and submitted; outcome not recorded. ## How I work Six habits, each with a file you can open. 1. I write the spec before the code, and get it approved. Proof: PostgreSQL fleet monitoring spec approved 23 July 2026, before the build. Thane Municipal deployment from a feasibility study approved 21 August 2026. Telematics design dated 16 September, revised 17 September 2026. 2. I record reversals instead of hiding them. Proof: A Go and Cloudflare telematics design was completed, reviewed, then discarded for an all-Rust single-VPS design; the 34 fixture packets and review findings carried forward. 3. I cap spend before the first real request. Proof: Vertex OCR runs at most 10 instances so a bulk upload cannot run unbounded paid calls; every invoice records its token cost in USD and INR. AgentReady has a daily run budget and payload caps. The desktop automation agent ships with a cost table ($0 idle, about $6 a month at 10 tasks a day). 4. A human signs off by default. Proof: Supplier-coordination agents route drafts through Teams Adaptive Cards for approval (designed and prototyped). QuoteDesk drafts a margin-loaded quote for a human to send. Proofline’s auditor accepts or rejects each mapping and that tunes the mapper. CD Infoware generates draft invoices, not final ones. 5. Docs ship with a CI check, or they rot. Proof: A CI job fails the CD Infoware build if the 12-chapter docs book’s code citations drift. The same docs-citation workflow runs on the react-native-fused-call package. 6. I verify what the model says instead of trusting it. Proof: Proofline’s verifier re-fetches every source and compares SHA-256 so hallucinated evidence is rejected. provenance-grade labels each extracted field verified, asserted or assumed. AgentReady records CAPTCHAs and logins as failures, never bypasses them, and labels its synthetic rows. ## Where I have shipped The stack changes; the shape does not: one interface per provider, explicit failure modes, a reconciliation path, a spend cap; I have built that shape for payments, travel suppliers and now LLM tools. - Payments (2018 to 2023): A Django payment gateway with processor failover, then a fintech marketplace wired to 30+ providers at 20,000+ transactions a day, leading a team of five. - Banking (2019 to 2020): 5 of 10 legacy Struts apps moved to Spring Boot and Angular in 10 months against an 18-month plan. - Travel (2023 to now): VIA Rail, JonView and Rocky Mountaineer integrations, Salesforce sync, Flywire payments, and a Postgres fleet monitor I designed, built and documented alone. - Field operations and AI (2026): Invoice OCR, a nightly meter scrape that drafts GST invoices in Tally format, a Rust telematics design for Teltonika and AIS-140 devices, and hackathon agents on ADK, MCP and WebMCP. Open source: an Android geolocation module with 109 Robolectric and 16 Jest tests, dogfooded in a client app. Hackathon builds are where I test new tooling on a deadline; client builds are where it earns its keep; both are on this site, labelled. ## Contact I am looking for a forward deployed engineer role at an AI-native company. Remote from Mumbai on North American hours. 8+ years integrating messy business systems; now integrating AI agents into them. Email: manasnikam24@gmail.com. GitHub: https://github.com/manasnikam. LinkedIn: https://www.linkedin.com/in/manasnikam. Resume PDF: https://manasnikam.dev/resume/Manas-Nikam-Resume.pdf.