Forward-Deployed AI Data Engineer · Boston / Remote

The deployment layer.

You wrote a job for the person who makes the promise real inside data that was never prepared for an agent. Models keep getting better. They still guess without the definitions a domain expert carries and the way a company actually uses its own tables. That is the only work I have done for twenty years: sit in the room, map the unofficial definition, make the pipeline honor it, leave a path the next engineer can run without me on Slack.

Edisyl’s own line: the technology works; what you are building now is the enterprise motion around it. I am that motion — Informatica’s parachute closer, 19 of 20 crisis accounts recovered, then the same playbook on a private RTX 4070 agent stack that measures. Forge, Lattice, and Stratum are names I have not worn. They describe work I have already done: agents with a bounded surface, a fleet that is traceable, a knowledge layer with owners. Year one success in your words — production implementations, playbooks from the field, clients asking for me by name, trusted to go in alone — is a year I have already spent, several times.

Open the brief IPS playbook Read the letter

Jeremy Panasuk · New Auburn, WI · the closer they send in alone

01 — What you believe

Models are not the constraint. Context is.

Your site says it plainly: agents fail expert work because they are missing the definitions a domain expert carries, plus the way a company actually uses its own data. I have sat in those rooms — with CTOs, with warehouse teams, with CRMs that lie — and left something running.

01 Capability

The model will still guess.

You wrote that a model can pass a board exam and still not know which of three definitions of an active patient the commercial team uses. I have lived that sentence on warehouses, not in a blog post. Dawn Foods: Type 2 slowly changing dimensions so “the customer” stopped meaning three different keys depending on who you asked. PepsiCo: NA/LATAM v9→v10 with zero data loss because the cutover honored the real grain, not the slide. Ameriprise: IDQ, Enterprise Data Catalog, Axon, more than 200 critical data elements held 24×7. The work is mapping the real definition, then making the pipeline honor it. Until that exists, the agent is guessing with better grammar.

02 Knowledge

Most of what matters is not in the table.

Regulations, naming, the workaround the night operator invented in 2011. Ameriprise EDG — IDQ, EDC, Axon. Smith & Nephew: FDA-validated Informatica environment, licensing and resource contention mitigated, best practices pushed in an org that was indifferent to app deployment, $600k licensing. Navy Federal: naming standards, services configuration, go-live / business-acceptance tests written down. I treat undocumented knowledge as a system of record, not a Slack thread. If it is not on the cover of the Environment Summary, it will not survive the next hire.

03 Ownership

Context has a maintenance cost.

Playbooks, not heroics. Mercedes-Benz: longest workflows +200%, performance-tuning documentation adopted by TCS so the next wave did not need me in the chair. Navy Federal testing strategy. Micron: resourcing plan 2,080 → 7,000 hours across PowerExchange, PowerCenter, Fast Clone, and Big Data, on time, extensions, zero downtime. About fifteen percent of IPS time was writing statements of work with regional managers — full contract wording, resource levels, estimates and assumptions. If the next engineer cannot run it, I did not finish. Ownership is the maintenance cost you actually paid.

04 Durability

The layer should outlive the model.

Informatica v7 → v10 as an upgrade specialty. PepsiCo v9 → v10 with zero data loss. Harvard Version 9 off Solaris 8.1.1, resource plan through post-production hypercare. Kohl’s: v9 Grid/HA capacity and upgrade, target architecture on commodity (Cisco or IBM), cutover off AIX, current volumes plus a three-year growth curve, briefed through Senior VPs, about $2 million in software, proofs of concept on their data when vendor VMs were not enough. I do not build a clever demo that dies when the vendor changes. I build the motion you can hand to the account.

02 — Knowledge pack

Two halves. Built together. Owned by the client.

You sell Domain knowledge + Proprietary knowledge on a semantic layer. I have been assembling that pack inside live accounts since 2005 — without calling it that.

Domain knowledge

Outside the walls

Insurance quote runtime — West Bend Mutual, live SOAP/XML path cut in half. Milwaukee insurer: three-month pile-up closed in two days on one buried IBM DB2 driver.
Banking / capital markets — BNY Mellon Grid/HA ($2.3M+ with CEO Sohaib Abbasi in the room). JPMorgan Chase DW/CoE roadmaps: we reached the final two proposals; we did not close; I will not write $10M as a win. Navy Federal 3,000-hour V1 with Informatica RMO plus their HR.
Wealth / governance — Ameriprise IDQ, Enterprise Data Catalog, Axon.
Healthcare / PBM — Prime Therapeutics; MOBE Redshift + SQL Server at 99.99%.
Manufacturing / CPG — Dawn Foods AS/400 CDC, PepsiCo NA/LATAM, Micron fab sites, Cargill CFO Award.

Proprietary knowledge

How this company actually runs

CRMs, warehouses, journals, SOAP, mainframe CDC, GitLab CI — the unprepared surface your agents have to survive.
I do not ask for a clean schema first. I find the bottleneck under the symptom.
Primary technical point of contact after go-live. Clients asked for me by name. Informatica IPS: “rock star.” MF Global infrastructure: “one of the best Informatica has to offer.”
Last 24 months: private inference lab, bounded MCP, production WhatsApp agents — same discipline, new substrate.

03 — Three components. One operator.

Forge, Lattice, Stratum — already in my hands.

I have not been on your payroll. I have been doing the job those three names describe.

Agent framework

Forge

Where agents are built and equipped. Tools, not raw dumps. Quasi-deterministic because they call surfaces with a boundary.

Mine: Four isolated Python agents on local Ollama (RTX 4070 / WSL2): Systems names constraints, Architect writes a Pydantic JSON contract, Synthesis generates the payload, Validator checks schema and graph and self-heals until it passes. Each agent has its own venv — a crash in one does not take the others down. OpenClaw agents with 10+ skills. Dockroot-MCP so agents see Docker/network without host-root. Tools, not raw dumps. Quasi-deterministic because the surface has a boundary.

Fleet orchestration

Lattice

Route the work. Track the dependency tree. Make every action traceable.

Mine: World-first Informatica Grid/HA + DR, 20+ nodes, four regions. local_grid_suite: SQLite benches, GPU routing, CLI + PDF. Measured ~27× local decode after routing vs CPU baseline.

Semantic intelligence

Stratum

The knowledge layer. Vocabulary in the systems vs vocabulary in the heads. Reconcile them or the agent guesses.

Mine: Ameriprise data governance. Type 2 warehouse at Dawn. FDA-validated platform at Smith & Nephew. I treat semantics as production, not a prompt.

04 — Your brief, answered

Every line in the job. Evidence, not adjectives.

Open any row. This is how I sit in a room with a CTO without making them feel stupid about their data.

The role

Embed inside client environments. Make agents work against unprepared data.

That was Informatica IPS / SEAL: parachute into burning Fortune-100 platforms, find the bottleneck under the symptom, leave a system other people can run. 20 troubled accounts entered, 19 recovered. Same motion now on local AI — measure, bound, hand off.

Onboarding

Discovery through production deployment

Navy Federal: 3,000-hour plan from architecture through go-live training. Harvard: v9 upgrade and Solaris offload through post-production support. Micron: 2,080 → 7,000 hours across PowerExchange, Grid/HA, Fast Clone, Big Data; zero downtime; Consultant of the Year runner-up (2016).

Connectors

Data connectors, pipelines, agent workflows in the client’s box

PowerExchange CDC on AS/400 remote journals (Dawn). DB2 on mainframe and AS/400 (Navy Federal). Talend 8 + GitLab CI/CD (MOBE). SOAP/XML realtime (West Bend). OpenClaw production agents + Dockroot-MCP. I write the integration. I do not throw it over a wall.

POC

Primary technical point of contact post-deployment

Clients asked for me by name. Dawn: “His knowledge not only of the Informatica suite but also at the operating system level has been truly invaluable.” IPS sales: “He’s a rock star out at Dawn. The customer loves him and the sales team has been leaning on him hard.”

Field → eng

Surface product gaps, failure modes, recurring patterns

That is how Grid/HA got sold at BNY Mellon — with senior VPs, support, and CEO Sohaib Abbasi in the room. Kohl’s capacity/upgrade roadmap to S-VPs. JPMC CoE evaluations. I do not file a ticket and disappear. I bring the pattern back so the next account is cheaper.

Playbooks

Implementation playbooks so the next one goes faster

Mercedes: +200% on the longest workflows; tuning documentation adopted by TCS. Navy Federal “Informatica Technical Stack” testing strategy. MF Global DR process for the entire platform, then the U.S. DR test. The IPS artifacts are public: Environment Summary, Upgrade Roadmap v3, Phase review, v10 test deck — Farmers Phase 2, Mercedes-Benz Financial 20–30%, Northwestern Mutual P1, Waddell & Reed v9→v10. Read the FDE playbook. Playbooks are how 19/20 stays 19/20.

Pre-sale

Partner on scoping, discovery, and proof-of-concept builds

Toyota Engine Management go-live with regional managers and sales — all consultants 5/5, $300k services + $500k licensing. Smith & Nephew $600k licensing on an FDA-validated platform. Kohl’s: v9 Grid/HA capacity, AIX off to commodity, three-year growth, Senior VPs, ~$2M software, POCs on their data when vendor VMs were not enough. JPMorgan Chase: final two proposals on an EDW / CoE roadmap — we did not close; I will not write $10M as a win. Volkswagen: complete Grid/HA in two weeks, day early, four consultants shadowing. I can sit with your Enterprise Data Strategist and CEO because I have sat with theirs.

SQL / Python

Think in queries. Python. APIs. DuckDB-class fluency.

Redshift and SQL Server pipelines at 99.99% (MOBE). Type 2 warehouse. IDQ rules. local_grid_suite is Python telemetry: discovery, decode/prefill benches, SQLite logs, CLI reporting. I do not look up how to join. I look up why this join is lying.

Agents

Hands-on AI agent workflows. Know where LLMs break on real data.

Private NVIDIA RTX 4070 lab: TabbyAPI/EXL3, Ollama, Open WebUI, GPU passthrough. Published: huggingface.co/jpanasuk/tabby-tavern-stack. Multi-provider routing with cloud failover. Agents get a bounded surface. Unbounded context is how they hallucinate a customer into existence.

Unstructured

No schema, no labels, no consistent format — did not flinch.

AS/400 journals. SOAP that never matched the WSDL. Four-region Grid with CIFS → GFS live. Before any of that: Secret Key Class B fireworks shooter and Fireworks Forever inventory lead, 1996–2004. Live show, $800k seasonal inventory, no second take. That is the nervous system this job needs.

Output

Bias toward output. Right results over elegant code.

Warehouse load 8 hours → under 1. Longest Mercedes workflows +200%. Real-time quote path −50%. ~27× local decode after GPU routing. The measure is not the PR. The measure is whether the thing ran tonight.

Opinion

Most AI deployments fail on data, not model — and I have proved it.

That is the whole career. Informatica did not fail because the engine was dumb. Accounts failed because the data, the runtime, the ops, and the politics were not ready. I recovered 19 of 20. Then I proved the same thesis on a single GPU with benches in SQLite instead of a slide.

Bonus

Palantir-style FDE. Financial services. Insurance. Blockchain: honest.

Informatica IPS was the consultative forward-deployed model — embed, close, playbook, repeat. Financial services and insurance are not a stretch: BNY Mellon, JPMorgan Chase, Navy Federal, Ameriprise, West Bend Mutual, Prime Therapeutics.

Blockchain / DeFi / 20+ chains / 700M wallets: that is your origin, not mine. I will not fake a chain analytics resume. Your next buyers are sitting on CRMs, warehouses, email, and document dirt. That is my origin. That is why you need me now.

05 — Year one, already spent

You described success. I have the receipts.

19/20

Crisis accounts

Entered burning. Left running. The team sent me alone.

8h → 1h

Dawn Foods

PowerExchange CDC on AS/400. Featured on informatica.com.

$2.3M+

BNY Mellon

Grid/HA with senior VPs and CEO Sohaib Abbasi.

+200%

Mercedes-Benz

Longest workflows. Playbook adopted by TCS.

20+ nodes

MF Global

World-first Grid/HA + DR. India, UK, New York, Chicago.

~27×

Local lab

Decode after GPU routing. SQLite benches. Not a lucky run.

Phase 2

Farmers

Guidewire XML. 30+ repos called as domain risk. Skill-up named in the review.

20–30%

Mercedes-Benz Financial

Cache in RAM, not more boxes. Cheat sheet left with the team.

Northwestern Mutual

Domain best practice. GCS escalation. QA twin of production.

v9→v10

Waddell & Reed

Measured I/O wait. UNC is not HA. In-place vs parallel vs clone.

06 — Asked for by name

Loyalty is the metric. Not satisfaction.

The line I have carried since the fireworks floor, later the Informatica stage: customer satisfaction is worthless; customer loyalty is priceless. Satisfaction is a survey they forget. Loyalty is the warehouse load that went from eight hours to under one and stayed there because they used it — and the BI lead who told a competitor no over lunch. Read the talk.

“Jeremy has been a pleasure and an asset to work with. His knowledge not only of the Informatica suite but also at the operating system level has been truly invaluable.”

Dawn Foods · client

“He’s a rock star out at Dawn. The customer loves him and the sales team has been leaning on him hard.”

Informatica IPS · sales

“It is the belief of the infrastructure staff that you have sent one of the best that Informatica has to offer in Jeremy.”

MF Global · infrastructure

“I trusted that Jeremy would deliver, with quality, on the project he was assigned. Jeremy has exceeded the expectations at someone of his grade level.”

Prime Therapeutics · client

07 — Letter

To the CTO and the Enterprise Data Strategist.

COVER.txt

Jeremy Panasuk
New Auburn, WI · 715-970-0204 · jpanasuk@gmail.com
github.com/jpanasuk-netizen · huggingface.co/jpanasuk · linkedin.com/in/jeremy-p-34203322/

Hiring team — Edisyl

You are hiring a Forward-Deployed AI Data Engineer to embed in client environments and make Forge, Lattice, and Stratum work against data that was never prepared for an agent. Every engagement has to end with something measurable. You want someone the client asks for by name.

That is my career, named twenty years late.

Informatica used me as the parachute. I recovered 19 of 20 crisis accounts. Dawn Foods warehouse loads went from eight hours to under one. Mercedes’ longest workflows moved +200%, and the playbook stayed with the team. BNY Mellon’s Grid/HA close sat north of $2.3 million with the CEO in the room. MF Global got one of the first Grid/HA plus DR environments in the world. I have sat in insurance, banking, wealth, healthcare, CPG, and fab — West Bend Mutual, Navy Federal, Ameriprise, Prime Therapeutics, PepsiCo, Micron, Farmers, Northwestern Mutual, Waddell & Reed, Mercedes-Benz Financial. The written operating path is public: the playbook.

You came out of blockchain data infrastructure. I did not. I will not pretend I resolved 700 million wallets. What I have resolved is the other mess: CRMs, AS/400 journals, SOAP, Teradata, Redshift, and the unofficial definition of a customer that lives in one director’s head. That is the environment your enterprise motion now has to survive.

Last 24 months I ran the same discipline on a private RTX 4070: a published local stack, measured GPU routing (~27× decode vs CPU baseline), bounded MCP so agents see Docker without host-root, and production WhatsApp agents with real skills. The thesis is the one on your site. Models keep getting better. They still guess without a knowledge layer that has owners.

Year one success, in your words: multiple implementations in production, playbooks from the field, clients asking for me by name, trusted to go in alone. I can start that clock on day one because I have already spent that year, several times.

I am remote from Wisconsin. High-autonomy teams only. If you want the person who makes the promise real, I am available.

Jeremy Panasuk