Internal AI software for workflows where guessing costs money.
I build internal AI workflow software for telecom and quote-heavy sales teams: quote builders, lifecycle/EOL knowledgebases, research agents, and review workflows. The model drafts; trusted data, deterministic code, citations, tests, and human approval control what ships.
Best fit: telecom, quote-heavy sales, and revenue/document teams where a wrong answer costs margin, compliance, or a customer.
- Built like production software — every build ships with tests, evals, and deployment gates.
- Safe by design — the model is structurally unable to invent a price, a spec, or a policy.
- Built by an operator — 15+ years taking revenue from $0 to $100M+, now building the machine behind the motion.
Start with an AI Workflow Teardown → one workflow mapped, scored, risk-tiered, and scoped — ending in a build / no-build call.
Measured per workflow on shipped systems — sources, methods, and limits are in each case study. Results, not guarantees.
Clarity, turned into software that runs.
PD Insights builds internal AI systems for the workflows where a wrong answer costs money — quoting, lifecycle decisions, research, and revenue ops.
I'm a revenue operator who also writes the Python, designs the schema, and gates the deploys — so the AI is pointed at pipeline, margin, and risk, not novelty.
Six systems I'm asked to build most
Each is delivered as working software — scoped, built, measured, and handed over with the tests that prove it.
Agentic workflows
Turn calls, voice memos, and unstructured input into committed, structured records — with rep-approval gates, not blind automation.
Knowledge systems (RAG)
Source-of-truth retrieval that stops the same question being asked forty times. Answers are designed to cite source material; unsupported gaps are disclosed instead of filled in.
Sales-cycle compression
Intake-to-quote workflows for complex catalogs — deterministic pricing, review flags, internal economics suppressed. Weeks become hours.
Lifecycle & EOL planning
Inventories turned into prioritized replacement pipelines — surfacing end-of-life risk before revenue walks out the door.
Revenue & outbound automation
Find, enrich, sequence, and report on target accounts at scale — CRM and Microsoft 365 wired in, admin handed to background agents.
Local & private deployment
On-prem-capable models tuned for latency and cost, so when required, sensitive legal, financial, and customer data can stay inside your environment.
Case studies
Seven deeper write-ups are linked below, with compact snapshots showing the range of systems shipped. Metrics are drawn from the linked case studies and should be read with the limits noted in each case study.
Quote Bot Hardening
Took a local-first Zulip quoting assistant from a near-failing code audit to a tested, eval-gated tool — reproducible builds, deterministic quote math, and a human-review gate, with the model structurally unable to invent a price.
"Pete got fast models running on our own hardware that are private enough for our most sensitive data — we're not sending anything to the cloud anymore."
— Jeff D. · Independent Telecom Consultant
Router Lifecycle Knowledgebase
Turns a twelve-tab device workbook into cited, end-of-life replacement answers in seconds — across 1,024 current and EOL devices, with every claim sourced or dropped, never guessed.
"Instead of digging through a twelve-tab spreadsheet, my team asks one question and gets a cited answer on what's end-of-life and what replaces it. It's the first thing they open now."
— Ken A. · Verizon Client Partner
Pricing & Quote Intelligence
A unified pricing and deal-intelligence engine — 4,600+ SKUs across 7 vendors compiled into one deterministic catalog that flags end-of-life hardware and surfaces real deal economics before a quote goes out.
"Quote prep went from a couple of hours to under ten minutes — and it catches discontinued hardware before it ever reaches a customer."
— Corey H., Sales Engineering Lead · Verizon Elite Partner
POTS Agentic Workflow
Turns a 60-second voice memo into a logged, researched, priced opportunity in under 30 seconds — saving reps 3–5 hours a week of after-call admin, behind a rep-approval gate.
"My reps record a 60-second voice memo and the opportunity is in the pipeline before they've left the parking lot. We've stopped losing deals to forgotten follow-ups."
— Ken D., VP of Sales · Telecom Agency
One Talk Call-Flow Designer
A design-and-spec tool for Verizon One Talk call flows — freeform engineer notes become a schema-validated deployment design, rendered as a visual map, a provisioning spec, and a plain-English customer summary. Every deployment designed the same way: as validated data, not a drawing.
Read the full case study →
Cellular Site Survey
A mobile-first, offline field-survey app that walks a technician through a site — cellular signal readings, POTS-line inventories, and photos — for cellular-internet, POTS-replacement, Starlink, and One Talk jobs. A deterministic on-device engine scores the readings; the AI narrates the recommendation but never makes it.
Read the full case study →More systems I've shipped
CRM Platform with AI & Microsoft Integration
Internal CRM tooling with AI enrichment wired into Microsoft 365 — find, enrich, sequence, and report on target accounts inside the Outlook and Excel tools teams already use, so a few people can run outbound at a scale that used to need a team.
Built for a cellular-router manufacturer.Dual Legal Analysis Tool
An adversarial "analyst + independent critic" workflow that pressure-tests every legal finding against an evidence rule before it reaches an attorney — surfacing weak or unsupported conclusions early, with a full audit trail and human sign-off on every call.
Built for a medical-malpractice law firm.POTS Tracker
Turns a sprawl of carrier bills and analog lines into one normalized inventory, a prioritized migration plan, and a live replacement tracker — so a portfolio that used to take weeks to quote gets mapped, costed, and ready to act on in hours, with no line slipping through the cracks.
Built for a POTS-replacement hardware manufacturer.Financial Analysis (Local-Model)
Parses statements and bills and computes deal-level economics on local models, so confidential financials can stay inside the building — turning what was a full day of analyst spreadsheet work into minutes of cited, exact-math output the finance team can stand behind.
Built for a major legal firm.Customer-Owned One Talk Approval Kiosk
A production Android approval kiosk shipped in ~5 days — a first Android build delivered via supervised AI engineering: 168 tests written before the code, encrypted token storage, a tamper-evident audit chain, and a human owning every boundary.
Single production device; soak window in progress. Case study →POTS Shutdown Tracker
Continuously tracks carrier copper and POTS shutdown and service-restriction notices by area, turning scattered regulatory filings into a plain-language risk read — so sites get migrated ahead of forced disconnections instead of scrambling after a line goes dead.
Built for a multi-site retailer.Have one messy workflow worth fixing?
Bring the workflow, the data sources, and the failure modes. I'll help map what can be automated safely and what should stay human-reviewed.
Start the 2-minute fit check →Assess → Prioritize → Design → Build → Tune
A repeatable path from "we have a messy workflow" to "it runs in production and we can prove it."
Assess
Map the workflow and find where time, margin, and risk leak.
Prioritize
Pick the one high-value workflow worth shipping first.
Design
Architecture, data contracts, safety boundaries, success metrics.
Build
Test-first implementation with evals and CI gates on every change.
Tune
Measure, fix the lowest-scoring behaviors, re-run, hand over.
Start with the right level of work
Discovery decides; building is separate. The fit call is free.
AI Workflow Teardown
Workflow map, hard-value ROI, agent-readiness score, risk tier, and a build / no-build recommendation. Build is separate.
Agent-Ready Workflow Audit
Readiness scoring, risk tiers, governance controls, and a ranked build roadmap.
Build Sprint
Working software with tests, evals, review gates, deployment, and handoff.
AI Ops Retainer
Monitoring, test runs, model/tool updates, fixes, and a managed change queue.
AI that can show its work
Every PD Insights system is built around approved data, clear tool boundaries, cited sources, exact calculations where math matters, test cases, logs, and named human approval before high-impact output leaves the workflow.
Approved data
What data may enter the workflow, which tool and account may handle it, and what must be redacted or excluded — decided up front.
Cited answers
Facts and recommendations point back to approved sources. Unsupported claims are flagged or refused, not filled in.
Human approval
Quotes, external messages, pricing, and legal language stop for a named reviewer before they leave the workflow.
Tests & records
Known cases, failure cases, logs, and review decisions create evidence that the workflow behaves as intended.
PD Insights provides operating, workflow, and technical controls — not legal advice, a formal audit, or certification. Regulated or high-risk uses should be reviewed by qualified legal, privacy, and security professionals.
How it ships
Hosted or fully private, audited end to end, and integrated with the systems you already run — here's the toolchain behind it.
Where it runs
How it integrates
How it's handed over
Tools under the hood: Claude · GPT · OpenAI Agents SDK · Hugging Face · Ollama · Python · FastAPI · React · Neon Postgres · Chroma · eval harnesses · CI/CD.
A revenue operator who can also build the machine.
Over 15+ years I've grown businesses from $0 to $100M+, scaled teams past 50, and built revenue lines from near-zero to multi-million-dollar scale at ORBCOMM and CTS Mobility — where I helped move Fixed Wireless from roughly 5% to 90%+ of activations — and co-founded and sold Venture Mobile.
Today, through PD Insights, I work across Madison AI and Masters Telecom — shaping go-to-market and building the AI behind it: agentic workflows, RAG knowledge bases, quoting engines, and revenue automation. My edge is rare: I can shape the plan, open the door, win the business, and build the machine behind it.
MS, Engineering Management
Quick fit check
Tell me about the workflow. If it's a fit, we book a 30-minute fit call from here — if it's not, I'll tell you honestly what to fix first. One boundary: design and build work starts after a Teardown, not on the call.
Prefer to talk first? Book a 30-min fit call →
Let's map one high-value workflow.
If you've got a messy, manual process that should be running as software, that's exactly the kind of thing I turn into clarity. Book a 30-min fit call — no deck required.