DyPol / v0.4 / Private beta

See what your engineering team is actually doing.

DyPol is an AI analyst that reads your GitHub data and source code, then answers any question you have about your engineering org — from "what did we ship last week" to "why is the billing PR taking so long."

Read-only GitHub App·SOC 2 in progress·No data resold, ever
ASK DYPOL
live
›

Aanya merged #1247 · auth/oauth-pkce on Tuesday and is currently deep in #1259 · billing/stripe-webhook-retries — she's pushed 14 commits since Monday but hasn't requested review yet.

One branch looks stuck: aanya/feat-usage-meter — last push 9 days ago, no PR opened. Want me to draft a nudge or summarize the diff?

02 sources · git + sql
⌘K to ask
Real screenshotLatency 1.2s
01 / The shift

Every other tool gives you a dashboard.
We give you an analyst.

Dashboards
last 7 days
Cycle time
3.4d
▲ 0.6 vs prev
Deploy freq
12.1/wk
▼ 1.4 vs prev
PR throughput
47
▲ 8 vs prev
Review latency
11h
▲ 2h vs prev
// You: now what?
// Dashboard: …
DyPol
a conversation, not a chart
Founder

Why is deploy frequency down this week?

DyPol

Two reasons. First, the staging cluster was down Tues–Wed (incident #84) and four PRs that were ready couldn't ship. Second, billing-svc has had only one reviewer this week because Marcus is on PTO — that repo accounts for ~40% of normal volume. It should normalize Monday.

Founder

Show me the four PRs.

DyPol

#1234 #1238 #1241 #1245 — all green, all approved, deploying now that staging is healthy. The first one is a 4-line copy fix; want me to start there?

Same data. Different interface.
02 / What you can ask

Real questions. Real answers. Real citations.

Three exchanges from a real customer environment. Names changed, repo names kept.

a.
People · staffing

› What is Aanya working on?

She's the primary author on #1247 · oauth-pkce (merged Tuesday) and is deep in #1259 · stripe-webhook-retries, with 14 commits since Monday but no review request yet.

Her branch aanya/feat-usage-meter looks stuck — last push 9 days ago, no PR. She also reviewed three of Marcus's PRs while he was on PTO.

b.
Throughput · diagnosis

› Why is the billing repo slow?

Two named reviewers carry 84% of billing-svc reviews (Marcus and Priya). Priya is on call this week, Marcus is on PTO — so review latency on that repo jumped from 6h to 38h.

Three PRs are blocked on review: #812 #814 #819. I can suggest a temporary CODEOWNERS reshuffle if you want.

c.
Code · across-repo audit

› Find every place we still call stripe.charge() — are they still active?

Five call sites remain. Three are reachable from production code paths; two are inside tests or feature-flagged branches that haven't shipped in 6+ months.

git grep -n "stripe.charge(" $(git rev-parse HEAD)5 results · 4 files
apps/billing/src/legacy/checkout.ts:142  await stripe.charge({ amount, source }) // ACTIVE — fallback path, hit ~30/day
apps/billing/src/legacy/checkout.ts:217  stripe.charge({ amount, source: token }) // ACTIVE — admin refund flow
apps/api/src/jobs/retry-charge.ts:58  return stripe.charge({ amount, source: src }) // ACTIVE — retry queue
apps/billing/test/legacy.spec.ts:33  stripe.charge({ amount: 100, source: 'tok' }) // test fixture
apps/api/src/experimental/dunning.ts:91  if (FLAG_DUNNING_V2_OFF) stripe.charge(…) // dead since 2024-11
03 / How the agent works

Eight tools. Any question.

DyPol's agent has access to a focused set of primitives — SQL on engineering metadata, full git command access on read-only repo clones, file reading at any commit, Python for analysis, chart rendering, and semantic search across PRs and issues. It composes these to answer whatever you ask.

run_sqlRead-only SQL across all your engineering metadata.01
shell_execAny git command on read-only clones of your repos.02
read_fileInspect any file at any commit.03
python_execCustom analysis in a sandboxed Python environment.04
make_chartRender charts inline with answers.05
describe_schemaSo the agent can author queries it didn't pre-know.06
semantic_searchFind PRs, issues, and comments by topic, not just keyword.07
summarizeCompress long context before reasoning over it.08
04 / Built for engineers, not against them

We measure pushed work. We tell engineers what we see. We never use "lines of code."

Scoring

Four signals: Impact, Quality, Collaboration, Consistency. Derived from review depth, revert rate, ownership of PRs that ship, and the work load engineers carry for one another. Never LOC. Never commit count.

Data access

Read-only GitHub App. Per-repo allowlist. Every agent action is in an audit log you can export. You delete your data with one click and we mean it — the deletion job runs within an hour. SOC 2 Type II in progress; report on request.

Engineer self-view

Every engineer can see exactly what DyPol sees about them. Same metrics, same trends, same prose. There is no "manager-only" view of a person. If a tool is going to look at someone's work, that person should be able to look back.

For

Two kinds of reader. One question.

Founders & CTOs

You stopped reading the dashboard six weeks ago.

DyPol replaces the ritual of "checking the metrics" with a one-line question. Ask once a week. Get a paragraph back. Forward it to the board.

"{quote from design partner}"
Founder · 22-person Series A
VPs of Engineering

One-on-ones, with the data already loaded.

Walk into reviews knowing what each engineer shipped, who they helped, and where they got stuck. Not as a scorecard — as a way to ask better questions.

"{quote from design partner}"
VP Eng · 60-person Series B
06 / Questions worth asking

Answers we'd want before installing this on our own repo.

Will my engineers see this as surveillance?
They will if you treat it that way. The product is built so engineers see exactly what you see — same metrics, same prose, same audit log. We've shipped it that way on purpose. Used as a 1:1 prep tool, it builds trust. Used as a leaderboard, it won't.
What data leaves our environment?
GitHub metadata and source needed to answer your questions. We never sell, share, or use your code to train foundation models. Embeddings stay in your tenant. On Growth and Enterprise plans, you can pin all inference to a region of your choice.
How is the score calculated?
Impact, Quality, Collaboration, Consistency — each on a 0–100 scale, derived from shipped work. The full formula is published, versioned, and runnable on your data. No opaque magic.
Can I self-host?
On Enterprise, yes — in a VPC of your choice on AWS or GCP. The agent and its tools run inside your perimeter; only billing telemetry leaves.
What if I want my data deleted?
One click in Settings. The deletion job runs within an hour and we send you a signed receipt with the row counts removed. Backups roll over within 30 days.

Install on GitHub.
See the dashboard in under ten minutes.