AWS Global Government Hackathon · Counter-UAS

AEGIS‑EDGE

Governable, resilient, cooperative counter‑drone defense — decided at the edge.
Machine‑speed decisions. A human authorizes every kill. The LLM never touches the trigger.
Syed “Wasi” Hussain · Team Wolfberg Per Aspera
[0:00–0:30] Say the thesis line verbatim, then advance. Don't sell features yet — sell the stance.
The problem · why we're different

Drones outpace humans. Two units can’t share a picture across a jammed link. And “AI that fires” can’t be accredited.

[0:30–1:15] Lead with the linking pain: “The military’s real problem isn’t detecting one drone — it’s getting two units, two convoys, to share one air picture across a jammed link and act on it together. That hand-off is manual and dies under EW. We make it automatic, and jam-resilient.” Then the contrarian close: everyone puts AI *in* the loop — we kept it out, which is the path to accreditation.
Live demo
LIVE DEMO
<1 msgate decision
<0.1%false‑positive vs friendly
45/45tests green
20drone saturation swarm
[1:15–7:30 · the majority of your time] Switch to the sim. Controls: R = radar down · J = jammer · Case 2 = auto DDIL cycle · Space = pause to explain · click a track for the AI threat‑call, then 💥 Destroy. End on beat 5: press R (threats go TRACK LOST) → release → they reacquire and the system auto-defeats them. That's your closer.
Scenario 1 · normal ingress — Manual mode

Drones inbound. A human authorizes every shot.

ManualCombinedAutonomous
Scenario 2 · saturation — Autonomous mode

Too many, too fast — machine-speed defense inside a bounded envelope.

ManualCombinedAutonomous
Scenario 3 · contested — DDIL / jammed

What happens when we get jammed?

Case 1 · clearCase 2 · DDIL / jammed
Amazon Bedrock — AI-generated assessment report

Bedrock turns the collected data into a plain-English report.

AEGIS-EDGE · THREAT ASSESSMENT REPORT (sample output)
AO: Joint Base Cascade — North Gate · WCS: TIGHT · 2026-07-01 08:14Z · engine: Bedrock (Claude)

SUMMARY — 4 tracks held · 2 classified HOSTILE · 2 defeated · 0 fratricide · 0 friendly engaged.

HOSTILE-1 (0427) — quad thermal + commercial-UAS C2, inside the no-fly. Classified HIGH.
  ROE: Tank Column · SLAMRAAM (cheapest in-range). Operator-approved. DEFEATED T+41s.
HOSTILE-2 (0318) — low RCS, fiber-guided (no RF emitter). Classified HIGH.
  RF-defeat N/A → escalated to SHORAD. Operator-approved. DEFEATED T+58s.
FRIEND-1 (0192) — valid ADS-B / Remote ID → COOPERATIVE. Hard-blocked; never engaged.

COMMS — Case 2 (DDIL): 1 jamming event (6s). Edge-autonomous; all tracks reacquired on recovery.
LEDGER — 27 entries · Ed25519 hash-chain VERIFIED · S3 Object-Lock (WORM).

ASSESSMENT — Nominal defense. Both hostiles neutralized with the cheapest effective effectors;
zero risk to cooperative traffic. Recommend continued TIGHT posture; pre-authorize the AUTO envelope
if swarm density rises. Model ran off the kill chain — it summarizes; it never decided a shot.
Generated by Amazon Bedrock from the tracks + audit ledger, after the gate decided. Illustrative sample.
The data · where it comes from 30%

Data the military already has.

All fused by a real IMM tracker · ingested via SAPIENT / BSI Flex 335 · no new sensors.
Zero new data collection. We're a software layer on the air picture you already own.
[~40s, right after the demo] Credibility slide. Say: “We fuse four real sensor modalities the military already collects — RF from DroneRF, thermal from HIT-UAV, and cooperative ADS-B tracks from adsb.lol and OpenSky — discriminate friend from foe, share the picture between units with a Cursor-on-Target-style link, and validate against real Ukraine UAV-strike data. No new sensors, no new pipeline.” Honesty line: real modalities/data, synthetic *scenario*.
Architecture & integration 20%

Edge decides. Cloud explains & remembers.

Edge — Greengrass / Snowball

Deterministic gate + sensor/effector I/O. Runs offline: when the link and cloud are denied, it keeps making shoot / no‑shoot calls locally.

the “EDGE‑AUTONOMOUS” banner in the demo = this boundary, live.

Cloud — GovCloud IL5

Bedrock generates the assessment report from the collected data (async & off the kill chain) · S3 Object‑Lock WORM (signed ledger) · KMS · IAM.

expensive path is bounded & throttled — never on the trigger.
[7:30–8:15] Show the official AWS diagram (/aws-diagram.html). Point at the boundary: “DDIL resilience is that edge/cloud line made real” — Kinesis → Greengrass gate → Bedrock, keys never on device (IoT TES). Name SAPIENT for partner integration.
AWS reference architecture 20%
[part of 7:30–8:15] Trace the flow: sensors → Kinesis → Greengrass gate (edge, offline in DDIL) → human gate → effector; ledger → S3 Object‑Lock (WORM); Bedrock explains async. “Keys never touch the device — IoT Token Exchange.” Edge decides, cloud explains & remembers.
Partner stack · built on the sponsors

An enterprise‑grade platform, not a one‑off.

  • AWS — Greengrass + Snowball (edge), Bedrock, S3 WORM, KMS, IAM on GovCloud IL5. The backbone.
  • Anthropic · Claude — the threat‑assessment LLM via Bedrock. Live in the PoC — and deliberately off the kill chain.
  • Okta — operator identity + MFA + RBAC; enforces our two‑person authorization on every engagement.
  • Elastic — searchable track history + audit ledger; vector context for assessments.
  • Datadog — observability: the <1ms gate SLO, Bedrock cost/latency, ledger‑write health.
  • Rapid7 — secures the defender itself: MDR + exposure management on the C‑UAS platform.
  • Red Hat — hardened edge OS (RHEL / Device Edge), OpenShift + Ansible for the fleet.
  • Spectro Cloud — Palette manages the edge Kubernetes fleet across sites at scale.
  • GitLab — CI/CD + DevSecOps: runs the 45 tests, signs & ships the edge components.
  • IBM — watsonx.governance for model risk + hybrid‑cloud on mission‑critical gov workloads.
[~30s, during the architecture beat] Flash this on the diagram. Say plainly: “Claude via Bedrock is live today; the rest are the enterprise stack we integrate — identity via Okta enforces our two‑person rule, Datadog watches the gate SLO, GitLab ships the edge.” Don’t overclaim — say “integrate,” not “built with.”
Security & feasibility 15%

A deterministic, auditable kill chain is accreditable.

[8:15–8:50] Lead with “certifiable, therefore fieldable.” The cost line is your optimization story for criterion 5.
Path forward · exactly how we replicate & field it 10%

From this PoC to a fielded system

PHASE 0 · NOW

Governed PoC

Cooperative + DDIL loop, 45 tests, Bedrock assessment. Validated on real Kaggle drone telemetry at the edge.

PHASE 1

Real sensors/effectors

Wrap radar / RF / EO‑IR + effectors behind SAPIENT. Deploy the deterministic core as an AWS IoT Greengrass component on a Snowball Edge on‑site. Swap synthetic tracks for live SAPIENT detections.

PHASE 2

Cloud backplane

GovCloud IL5: Bedrock endpoint, S3 Object‑Lock for the signed ledger, KMS/IAM, CloudWatch telemetry.

PHASE 3

Range pilot + ATO

Use the auditable kill chain to drive accreditation; controlled live‑range test with real effectors.

PHASE 4

Federate & harden

Multi‑site Link‑16 federation, PACE comms fallback, bandwidth‑prioritized sharing, confidence‑based shot‑withhold, EW red‑team.

# replicate the PoC today
git clone github.com/wasihussain914/Aegis-Edge && npm i
npm test # 45/45 green — the governed loop is the test suite
npm run dev # live sim · bridge/bedrock.mjs = real Bedrock (SigV4, offline fallback)
# integration seam = the SAPIENT adapter; core is pure & deployable as a Greengrass component
[8:50–9:40] This slide answers “how do you actually move forward.” Emphasize SAPIENT as the single integration seam and Greengrass as the deploy target.
Cost & optimization 10%

The money is in sensors & effectors — we're the cheap keystone.

~$15–75/mo cloud · per site
FLATcost under a swarm
1–5%cloud share of TCO
$2.5B→$30.3Bmarket by 2035
[~20s · part of the 8:50–9:40 criterion-5 window, after the roadmap] One line: "The expensive part — the LLM — is off the hot path and throttled, so cost is flat under a swarm; at fleet scale our software is a few percent of TCO, on a market headed to $30 B." Full detail in COST-ANALYSIS.md.
Collaboration · team C2 console

One system, two views.

My 3D tactical sim

The air picture and the engagement — detect → classify → human-gated defeat, cooperative Link-16 fusion, DDIL resilience.

Berg's 2D operator console

The command-and-control layer — weapons-control posture (HOLD/TIGHT/FREE/AUTO), saturation → pre-authorized auto-defense, and the hash-chained audit ledger.

▶ Launch Berg's C-UAS operator console
Same governed spine — the deterministic gate + hash-chained ledger — seen from the tactical edge and the operator's C2.
Team Wolfberg Per Aspera · 2D console by Berg. Also federates with a partner multi-domain C2 (JADC2) — see the repo.
[at the end] Open Berg's console link and run it live: "Same governed spine, two vantages — my 3D tactical sim and Berg's operator C2. Weapons-control posture, saturation → auto-defense, the hash-chained ledger." Credit Berg.
Close
Machine speed. Human authority. The LLM off the trigger — and it keeps fighting when the comms are gone.
<1 msdeterministic gate
<0.1%FP vs friendly
45/45tests
DDILedge‑autonomous
Repo: github.com/wasihussain914/Aegis-Edge  ·  Live demo running now.
Syed “Wasi” Hussain · Team Wolfberg Per Aspera · Thank you.
[9:40–10:00] Restate the one‑liner, name the ask (pilot partner / next step), thank them. Leave the sim on screen.
1 / 7← → or Space to navigate · F fullscreen